Corrections & modifications

• Date: 2026-10-25
• Page: About, FAQ, How We Communicate, How We Use AI, the Claim Ledger, What We Don't Yet Know, Contribute and Objections (the promise); and about forty English pages and one blog post (the sentences)
• What was wrong: The site promised a source only for claims with a number, a date, a named study or a quotation. A pass over every English page found 129 factual sentences that linked no source on their page. Five said more than the page's sources: Mental Health credited "several large studies" with a finding its one linked study carries; Xenophobia said research "consistently" finds a link between online hate and offline violence, with one study on the page; Educators called Harmony Square a "Cambridge/Jigsaw" game; Advertising said fingerprinting survives private browsing and most ad blockers; Get Rich Quick said scammers reach "the most receptive people".
• What we changed: The promise now reads "every factual claim links to its source"; the site's own argument stays its own. Where an entry on the page carries a sentence, the sentence now links it; where none does, the sentence is narrowed to what a source says, or said as the site's own reading. The five overstatements now say what their sources say. Body Image no longer says that looksmaxxing culture promotes "starvemaxxing", which clinicians recognise as disordered eating, nor that looksmaxxing and bonesmashing can shade into body dysmorphic disorder and eating disorders: no source on the page carried either, and they come back only with one. Nineteen sentences were left as they were, with what would source them, in the review's report (docs/review/sourcing-promise-2026-10-25.md).
• Source: Internal review: one Fable read per English page (#175, card sourcing-promise-scope = a); the ICO's fingerprinting statement (19 December 2024), the FTC's Data Spotlight (April 2026), KVK's page on the stichting, Kelly et al. (2018).

• Date: 2026-10-24
• Page: Elections (/en/symptoms/elections/), How It Works (/en/attention-machine/how-it-works/), For Lawmakers (/en/for-you/lawmakers/)
• What was wrong: The three pages carried one observational study (Germano, Gómez and Sobbrio, 2026) on Facebook's 2018 ranking change and polarisation, and did not name the largest experiments on the question: the 2020 US election studies run with Meta, in which three months of a chronological feed, no reshares, or less like-minded content did not measurably change polarisation or political attitudes. How It Works also called the direction of effect "well-supported".
• What we changed: Elections names the three experiments and their limits (three months, one election, and Meta's post-election emergency ranking changes, which critics argue in a letter to Science changed the comparison feed); How It Works and Lawmakers name them in their contested sections. Four new Sources entries. The Brazilian Portuguese text follows in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings T1, E4 and M8; Guess et al. 2023 (Science 381, two papers); Nyhan et al. 2023 (Nature 620); Bagchi et al., Science eLetter, 26 September 2024. Card en-read-election-experiments = a.

• Date: 2026-10-24
• Page: Advertising & Surveillance (/en/attention-machine/advertising/), Elections (/en/symptoms/elections/), Timeline (/en/timeline/)
• What was wrong: Advertising said Cambridge Analytica used the harvested data "to build psychological profiles for political micro-targeting in the 2016 US presidential election. This is a confirmed fact". The cited sources confirm the harvesting and the transfer; which campaign the data served, and what the "profiles" amounted to, are disputed. Elections and the Timeline said the same more briefly.
• What we changed: The pages now say the company used the data in its work for US political campaigns in the 2016 cycle, and that the harvesting and the transfer are the confirmed facts. Whether the presidential election can be named again waits on a reading of the UK Information Commissioner's letter of 2 October 2020. The Brazilian Portuguese text follows in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings M17 and T14. Card en-read-cambridge-analytica = a.

• Date: 2026-10-24
• Page: State Actors (/en/attention-machine/state-actors/), Radicalisation Pipelines (/en/symptoms/radicalisation/)
• What was wrong: State Actors' section on European intelligence agencies was mostly about domestic radicalisation (MI5's terrorism arrests of children, the BfV's extremism-related crimes, the AIVD's jihadist caseload); only one item concerned a state actor.
• What we changed: State Actors keeps the state-linked item (a minor who mapped networks in The Hague for a Russian state-backed hacking group) under a sentence saying what the agencies' reports are evidence of; the domestic figures moved, with their sources, to a new section of Radicalisation Pipelines, "What the security services see". The Brazilian Portuguese text follows in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings M26 and E5. Card en-read-state-actors-agencies = a.

• Date: 2026-10-24
• Page: About (/en/about/), Journalism (/en/truth/journalism/), Science (/en/truth/science/), Symptoms (/en/symptoms/), Age Restrictions (/en/laws/age-restrictions/), Donate (/en/share/donate/)
• What was wrong: About called understanding the technique "the most effective defence", a ranking no study makes. Journalism said feed algorithms have "no correction mechanism", and Science that none of the systems shows its corrections or a record of changes; labels, reader notes and appeal routes exist. The Symptoms headline said the harms were "ignored for a decade"; the pages show harms known and not acted on, over more than a decade. Age Restrictions said "The direction is correct" right after calling the evidence thin. Donate said "Reader donations keep it running" and "This is currently the only income"; no donation had been received.
• What we changed: "The best-evidenced personal defence"; the ranking has "no correction mechanism of its own", and nothing records what it showed to whom; "for years not acted on"; age limits with design reform as "a lever worth testing"; donations as "what will keep it running" and "the only income the site will take". The Brazilian Portuguese text follows in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings E22, T7, T5, M29 and E20; the Q3 2026 transparency report. Cards en-read-key-line-3, en-read-symptoms-ignored, en-read-age-direction and en-read-donate-income = a.

• Date: 2026-10-23
• Page: Organisations (/en/share/organisations/), Donate (/en/share/donate/)
• What was wrong: Organisations said EUvsDisinfo "keeps the searchable record of documented cases". Its database is of pro-Kremlin disinformation, and it is a project of the East StratCom Task Force, part of the EU's diplomatic service, as its own About page says. Donate said registration "unlocks" Google Ad Grants and TechSoup discounts; registration makes the foundation eligible to apply, and each programme checks an applicant's non-profit status first.
• What we changed: Organisations now names the scope and the sponsor: an official project with one adversary in view. Donate says the foundation becomes eligible to apply, and that both programmes check status. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166, 23 October 2026), findings E19 and E21; euvsdisinfo.eu/about; Google for Nonprofits eligibility guidelines.

• Date: 2026-10-23
• Page: For Teens (/en/for-you/teens/), For Parents (/en/for-you/parents/), For Educators (/en/for-you/educators/), What to Do (/en/educate/what-to-do/), Watch & Listen (/en/educate/watch-listen/), Share This (/en/share/share/), Organisations, Tools & Sites (/en/educate/tools/)
• What was wrong: Teens said outrage and fear "get shared more widely than things that simply make you happy"; the study cited measured moral-emotional words in political messages and did not compare them with happy content. Parents, What to Do and Watch & Listen called watching Adolescence together "one of the most effective conversations available", What to Do called the Bad News game "more effective than any explanation", Educators and Organisations called prebunking "more effective" than debunking, and Watch & Listen called The Social Dilemma one of the most effective pieces of awareness-raising "in the public domain". No study compares any of these, and a Netflix title is not in the public domain. What to Do said "reading this site is itself a protective act"; the research tested games and videos, not this site.
• What we changed: Teens: messages with that moral charge travel further. The recommendations are now recommendations: a natural way to open the subject; a game that shows the techniques from the inside; prebunking that reaches further, because debunking arrives late and works claim by claim; the most widely seen documentary of its kind. What to Do says the interventions have been tested and this site has not. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings E9, E10, E12, E16 and E17.

• Date: 2026-10-23
• Page: For Parents (/en/for-you/parents/), Mental Health (/en/symptoms/mental-health/), Yes, but… (/en/objections/)
• What was wrong: All three said Facebook's internal research, disclosed by Frances Haugen, was "a different kind of evidence" from a disputed correlation, finding harm to teenage girls. Those were internal surveys, no stronger than the published studies; what they establish is that the company saw the signal and the product continued.
• What we changed: The three pages now say what the documents show: conduct, which no effect-size debate touches. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), finding E7.

• Date: 2026-10-23
• Page: Courts & Liability (/en/laws/courts/)
• What was wrong: The lead said the argument that sank the tobacco companies "is now being made — and won — against social media platforms"; the page itself says neither verdict is final. The four tobacco parallels were stated as the site's findings: that the platforms' knowledge "was kept from the public and from regulators", and that "children are a documented target demographic", with no source for either. Both are what the states and the families allege.
• What we changed: The lead now says the first two juries to hear the argument accepted it, in verdicts that are not yet final. The list is introduced as what the plaintiffs allege, and the concealment and targeting points quote the California Attorney General's release on the states' complaint. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166, 23 October 2026), finding M30.

• Date: 2026-10-23
• Page: Tracking & Targeting (/en/attention-machine/tracking-and-targeting/)
• What was wrong: It said IAB Europe is a joint controller of the consent string "subject to that court's checks", and that "the consent that RTB relies on has been found invalid". The Brussels Market Court ruled on 14 May 2025: it confirmed the joint controllership for the consent string and the €250,000 fine, and held that IAB Europe is not a controller of the processing inside the auction itself. No ruling found every consent given through the framework void; the framework was found not to comply.
• What we changed: The page now cites the ruling, through the Belgian Data Protection Authority's release, and says the consent mechanism "has been found not to comply with the law". The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), finding M19; Belgian Data Protection Authority, 14 May 2025.

• Date: 2026-10-23
• Page: Advertising (/en/attention-machine/advertising/), Tracking & Targeting, For Lawmakers (/en/for-you/lawmakers/), Search (/en/attention-machine/search/), Your Brain (/en/attention-machine/your-brain/)
• What was wrong: The ICCL's 376 and 747 daily broadcasts were given as measurements; they are the ICCL's 2022 estimate from industry data. Advertising said the 1999 concern about profiling was "regulated" before being set aside; the FTC closed the file without action. It said fingerprinting "cannot be defeated"; several browsers now blur it. Search said ads sit at the top of "most searches" and of "any commercial search", and that autocomplete "has repeatedly been found" to suggest extremist completions, on the strength of one 2016 article. Your Brain said threat content "consistently outperforms content that produces calm"; the study cited measured moral-emotional words in political messages, about 20% more spread per word.
• What we changed: "An estimated" and "the ICCL's 2022 estimate"; "investigated"; fingerprinting "survives" most countermeasures and "none makes you invisible"; ads on "the searches advertisers care about", "usually"; the 2016 article dated as one investigation, with its own argument stated as its own; "tends to spread further", with the study's figure. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings M9, M14, M15, M16, M22, M23 and E26.

• Date: 2026-10-23
• Page: AI (/en/attention-machine/ai/), The Attention Machine (/en/attention-machine/), Timeline (/en/timeline/), State Actors (/en/attention-machine/state-actors/)
• What was wrong: The AI page's lead, description and card said generative AI "makes" manipulation dramatically cheaper, and the Timeline's 2022 entry that it "makes" it cheap; the source the page cites is a forecast, and the page's own evidence section finds no conclusive effect yet. State Actors described the Meliorator software in its own voice, on the strength of the intelligence advisory alone, and said "polarisation is profitable for platforms", which the platforms dispute.
• What we changed: Generative AI "is expected to" make manipulation cheaper, and "is removing" the constraint. Meliorator is described as the advisory describes it, and the US Department of Justice's announcement of the same day — the seizure of two domain names and the search of 968 accounts on X, under court warrants — now corroborates it. "Polarising content is engaging, and engagement is what the platforms sell." The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings M24, M25 and M27; US Department of Justice, 9 July 2024.

• Date: 2026-10-23
• Page: Laws (/en/laws/), Timeline, Gamification (/en/attention-machine/gamification/) and its demo, Data Combination (/en/attention-machine/data-combination/), the DSA (/en/laws/dsa/), the GDPR (/en/laws/gdpr/)
• What was wrong: Laws and the Timeline called the GDPR the first privacy law "with penalties that bite"; the 1995 Directive carried penalties through national law, and the GDPR's difference is their scale. Gamification said variable rewards are "unpredictable enough that you cannot stop", and its demo called them "the most reliable way known to psychology" to keep a behaviour going; the studies compare a few basic schedules. Data Combination said "no company set out to build that profile" beside a data-broker card whose business is building it. The DSA page drew a verdict from a study of two countries and one platform that its own Deeper box calls a signal. The GDPR lead used "surveillance capitalism", a scholar's coinage, unattributed.
• What we changed: "Penalties on the scale of the companies it addressed"; "stopping is hard"; "of the basic reward schedules psychology has compared"; the broker named among those who build the profile; "a signal rather than a verdict"; "behavioural advertising". The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings M12, M13, M21, M28, M31 and M32.

• Date: 2026-10-23
• Page: Timeline (/en/timeline/)
• What was wrong: The 1996 entry said Section 230 "protects platforms, not users", and the paragraph under the timeline that it gave platforms protection "rather than giving users any". The law's operative sentence protects both: "No provider or user of an interactive computer service shall be treated as the publisher or speaker" of what someone else posted. What it does not do is give the person harmed a remedy against the service.
• What we changed: Both now say it shields the online service, the intermediary, rather than giving anyone harmed a remedy. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166, 23 October 2026), finding T22, checked against 47 U.S.C. § 230(c)(1).

• Date: 2026-10-23
• Page: Elections (/en/symptoms/elections/), Timeline (/en/timeline/), AI (/en/attention-machine/ai/)
• What was wrong: Facebook's own disclosure says the data of "up to 87 million people" may have been shared with Cambridge Analytica; Elections and the Timeline gave the ceiling as the count. Elections and AI called Călin Georgescu a "previously unknown" candidate; he was a little-known one, with a public record before 2024. Elections also said the declassified Romanian documents were "showing" the TikTok amplification, which states the intelligence finding as fact; the AI page has said "stating" since 19 October.
• What we changed: "Up to 87 million" on both pages; "little-known" on both; Elections now says the documents were "stating" that he had received the amplification. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings T13 and T14; Facebook, 4 April 2018.

• Date: 2026-10-23
• Page: The Manosphere (/en/symptoms/manosphere/)
• What was wrong: It said intelligence agencies "are treating this", the manosphere, "as a security concern", and gave MI5's figure for all terrorism investigations of under-18s and the AIVD and BfV reports on young right-wing extremists. None of those sources is about the manosphere. Only the Dutch forensic mental-health services, reported by NOS, are.
• What we changed: The security services' figures are now given as the wider pattern of young people radicalised largely online, of which the manosphere is one strand, and the forensic services are named as the source on the manosphere itself. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), finding T15.

• Date: 2026-10-23
• Page: Consumerism (/en/symptoms/consumerism/), Body Image (/en/symptoms/body-image/)
• What was wrong: Consumerism said the effect of psychographic targeting on purchases "is established by field experiment". The experiments it cites (Matz et al. 2017, over 3.5 million people reached) did not assign at random who saw which ad, and a published letter (Eckles, Gordon and Johnson, PNAS 2018) argues that this threatens their causal conclusion; the authors replied. Body Image's Deeper box said the "correlation is strong", above a meta-analysis that the page itself describes as finding a small but significant link.
• What we changed: Consumerism now says the claim has field evidence behind it, names the design limit and cites the letter and the reply. Body Image now says the correlation is consistent and the effect small. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings T10 and T12; Eckles, Gordon & Johnson (2018), PNAS 115(23), E5254–E5255.

• Date: 2026-10-23
• Page: What We Don't Yet Know (/en/symptoms/unknown/), AI (/en/attention-machine/ai/)
• What was wrong: Both said the commercial incentive structure of AI companions "is identical to social media". Many companion products are paid for by subscription, not advertising; the incentive to maximise return use is shared, the advertising model is not.
• What we changed: Both now say the incentive is close to social media's, even where the money comes from subscriptions. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings T19 and M24.

• Date: 2026-10-23
• Page: Science (/en/truth/science/), Get-Rich-Quick (/en/symptoms/get-rich-quick/), Xenophobia (/en/symptoms/xenophobia/), Mental Health (/en/symptoms/mental-health/), Yes, but… (/en/objections/)
• What was wrong: Science said the replication reforms "produced measurable improvement in research practices"; the study it relies on measured reported p-values, a signature of the change, as its own Deeper box says. Get-Rich-Quick called pig-butchering "the fastest-growing investment fraud", which no source on the page says. Xenophobia called the causal pathway from online hate to violence "real and documented" in general, where the study behind it covers one country and period. Mental Health called the word "addiction" "inaccurate" without qualification, where the study it cites found that most, not all, who use it of themselves do not meet clinical criteria. Yes, but… offered terrorism arrests of children as evidence that they are "not immune to the engagement algorithm"; the arrests show radicalisation, not what caused it.
• What we changed: "Measurable signs that research practice has changed"; "a fast-growing kind of investment fraud"; "a causal pathway has been documented, with rigour, in one country and period"; "usually inaccurate"; and on Yes, but…, the audit evidence on what the feed recommends to young male accounts, with the arrests kept as the far end and not as proof. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings T9, T17, T18, T21 and T23.

• Date: 2026-10-23
• Page: Courts & Liability (/en/laws/courts/), Section 230 (/en/laws/section-230/)
• What was wrong: Both pages said US courts were finding that Section 230 does not protect platforms over their own design choices, naming infinite scroll, autoplay and push notifications (Courts added notifications "calibrated to moments of maximum psychological vulnerability", which no source on the page carried). The ruling both pages cite, the federal order of 14 November 2023 in the youth-harm litigation, held the opposite for those features: claims over the feed's lack of a "beginning and end", the "use of algorithms to promote addictive engagement" and the timing of notifications were barred. What it let go forward were the claims over parental controls, age verification, barriers to deleting an account, filters, reporting tools and failure to warn. The Courts page's own timeline entry said so, against its lead.
• What we changed: Both pages now name the design claims the ruling allowed, and say that the feed's own features are the line still being drawn: barred in that ruling, while the Third Circuit let a suit over TikTok's recommendations go ahead in 2024. The order itself is now cited beside the law firm's summary. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166, 23 October 2026), finding M5, checked against the order (N.D. Cal., No. 4:22-md-03047-YGR, Dkt. 430).

• Date: 2026-10-23
• Page: AI (/en/attention-machine/ai/)
• What was wrong: It said of the annulled Romanian election that "the algorithmic amplification is confirmed". That rests on the declassified intelligence note alone (as the state news agency reported it), which the Constitutional Court accepted and constitutional-law scholars criticised. The site's rule is that intelligence reporting corroborates a claim and never carries one alone.
• What we changed: The page now says what the sources support claim by claim: the coordinated amplification rests on the declassified note, which the Court accepted and critics dispute; that influencers were paid to post content that ended up promoting the candidate is also documented outside the note, by a news investigation that drew on the tax authority; the role of generative AI is a further, thinner claim. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), finding M4.

• Date: 2026-10-23
• Page: How It Works (/en/attention-machine/how-it-works/)
• What was wrong: It said a recommender decides "based on one signal" and that accuracy and wellbeing "are simply not inputs to the model". The source it cites describes engagement as the core logic, with secondary inputs added on top, such as a score for the likelihood that a post is hate speech and a classifier trained on user surveys. It also gave the 64% of joins to extremist groups as a figure "on the platform", where the internal slide, as the page's own Deeper box says, studied political groups in Germany.
• What we changed: The recommender now decides "overwhelmingly, on one kind of signal", and other inputs are described as adjustments layered on top, not the quantity the system maximises. The 64% figure is now given for the German political groups the slide studied: one country, one year, the company's own measurement. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings M1 and M2.

• Date: 2026-10-23
• Page: Health Misinformation (/en/symptoms/health-misinformation/)
• What was wrong: Under "The cost", the worldwide measles figures for 2019 were followed by a clause tying the surge to countries where vaccination coverage slipped, which on a page about misinformation reads as misinformation's doing. The WHO release cited attributes the surge to "a failure to vaccinate children on time" and does not mention misinformation.
• What we changed: The page now gives the WHO's explanation and says that misinformation is one of the reasons people decline a vaccine, not the explanation for the global figure. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), finding T3.

• Date: 2026-10-23
• Page: Mental Health (/en/symptoms/mental-health/), Timeline (/en/timeline/)
• What was wrong: Mental Health stated flat that adolescent mental health began deteriorating around 2012, as smartphones spread, and marked the timing contested only two paragraphs later. The Timeline's 2012 entry said researchers "begin documenting" the decline that year; the decline was dated to around 2012 only years later, and that dating is the contested part.
• What we changed: Mental Health now attributes the timing to the measures Jonathan Haidt assembles and gives his critics' objection in the same sentence. The Timeline entry now says the decline was later dated to around 2012, mainly on US surveys, and that how widespread it was and what caused it are contested. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), finding T2.

• Date: 2026-10-23
• Page: Elections (/en/symptoms/elections/), How It Works (/en/attention-machine/how-it-works/), For Lawmakers (/en/for-you/lawmakers/), Young Adults (/en/for-you/young-adults/), Democracy (/en/truth/democracy/)
• What was wrong: One observational study (Germano, Gómez and Sobbrio, survey data from the US and Italy) became "measurable effects on political opinion at national scale" on Elections and "political consequences at national scale" on How It Works; Lawmakers stated it as fact, and Lawmakers and Elections listed it under "Confirmed". Young Adults spoke of "the political polarisation of the feed". Elections and Democracy said the system amplifies "false, inflammatory, and polarising content", where the evidence on the pages is about inflammatory and polarising content.
• What we changed: Each page now says it is one study of survey data, not an experiment, and that how far a single ranking change moves polarisation is not settled; the study moved from "Confirmed" to "Contested" on Lawmakers and Elections; Young Adults says "the polarised feed"; Elections and Democracy say "inflammatory and polarising content, true or false". The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings T1 and E4.

• Date: 2026-10-23
• Page: How We Communicate (/en/about/how-we-communicate/), Contribute (/en/share/contribute/), FAQ (/en/faq/), Yes, but… (/en/objections/)
• What was wrong: They said a translation goes live only after a native speaker has reviewed it. Since the Brazilian Portuguese edition was published while its volunteer reviewers are still reading it, that is true of every edition but one kind; How We Use AI and Translation Status already said so.
• What we changed: The four pages now state the exception: an edition published under volunteer review, which says so on every page. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), finding E6.

• Date: 2026-10-23
• Page: For Parents (/en/for-you/parents/), For Lawmakers (/en/for-you/lawmakers/), Teens (/en/for-you/teens/), What to Do (/en/educate/what-to-do/), Age Restrictions (/en/laws/age-restrictions/), Search (/en/attention-machine/search/), State Actors (/en/attention-machine/state-actors/)
• What was wrong: Parents offered "the primary cause" or "a significant contributing factor" as the only answers to a causal question that is itself contested; Lawmakers said the harms on this site are not contested at the level of whether they exist, which for adolescent mental health is exactly what is contested; Age Restrictions spoke of "the design features that cause harm". Parents, Teens and What to Do said notifications are timed to "the moment of maximum vulnerability", and Teens that the feed targets "the hour your willpower is lowest": design-intent claims no page sourced. Search turned one laboratory study into "Position controls belief". Lawmakers and State Actors said intelligence agencies now say "what researchers have been documenting", where the agencies say that radicalisation happens online and nothing about engagement ranking.
• What we changed: Each sentence now says what its sources carry: the causal question is marked contested where it is asked; mental health is named as the exception on Lawmakers; the harms are "attributed to" design; notifications are sent to bring the user back; position "moves belief — in that study, measurably"; and the agencies are quoted for what they say, which is narrower than the site's argument. The other editions follow in the same change.
• Source: Internal review: the second Fable read of the English argument (#166), findings E1, E2, E3, E5, M3 and M6.

• Date: 2026-10-22
• Page: Privacy (/en/about/privacy/)
• What was wrong: In the table of who processes what, the Cloudflare row said "Standard server logs, no personal data". A server log records each request's IP address, and an IP address can be personal data under EU law (CJEU, Breyer, C-582/14, 2016).
• What we changed: The row now says "Standard server logs, which include your IP address, kept by Cloudflare under its own policy (we do not receive them)": the site's own Worker logs have been off since 7 October 2026, and Cloudflare, as the host, keeps its logs under the privacy policy the row links. The Dutch, Romanian, Brazilian Portuguese, French and Spanish editions carry the same change.
• Source: Internal review: the Fable review of 10 October 2026 (§4), collected in #164; Decisions card en-privacy-server-logs = c.

• Date: 2026-10-22
• Page: For teens (/en/for-you/teens/)
• What was wrong: It introduced its list of help lines with "These services are free", but the list includes lines that are not, or that do not say so: Meri Trustline in India is a mobile number, and Ora de Net in Romania is not called free by its own site.
• What we changed: The sentence now says that most of these services are free and that each entry says what it costs where that is stated — the wording the Get Help page already uses for its own list of support lines. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow in the same change.
• Source: Internal review: the Fable translation review of 26 September 2026, collected in #164.

• Date: 2026-10-22
• Page: Share (/en/share/)
• What was wrong: The card for Things Worth Paying For said "Referral relationships disclosed", which suggests there are some. The page it links to says the site has no referral or affiliate income at all.
• What we changed: The card now says "No referral or affiliate income", as the page does. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow.
• Source: Internal review: the Fable review of the Romanian section 5 drafts, 10 October 2026, collected in #164.

• Date: 2026-10-22
• Page: Where to get help (/en/educate/get-help/), entry cybercrime.gov.in · 1930
• What was wrong: It said "anonymous reporting possible" for the portal as a whole. The portal offers anonymous reports only for crimes against women and children; a report of financial fraud cannot be anonymous.
• What we changed: The entry now says that a report about a woman or child can be anonymous and a financial-fraud report cannot, as the Dutch, Romanian and Hindi texts already did; the Brazilian Portuguese, French and Spanish texts follow. The same change removes a repeated "(free)" from the Offlimits entry (its line is already described as free and anonymous) and adds deepfakes to what Meri Trustline helps with, as its partner Aarambh India describes it and the Hindi text already said; for those two the Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow.
• Source: Internal review: the Fable translation review of 26 September 2026 and the Fable review of !247, 10 October 2026, collected in #164; the portal's own reporting page, as recorded in docs/research/IN.md.

• Date: 2026-10-22
• Page: AI & The Next Layer (/en/attention-machine/ai/)
• What was wrong: It said the 2024 European Parliament elections were "described by researchers as a 'testing ground'" for AI. The abstract of Farooq et al. (2025) says the elections "were regarded as" one: the authors report a view held before the vote, which their study then tests, not their own verdict.
• What we changed: The sentence now says the elections were "regarded as a 'testing ground'". The same change quotes Cornea's own words for those who defended Romania's annulment — "an expression of militant democracy" — on the AI and Elections pages, and gives the Timeline's 2007 entry in date order (the iPhone in June, Facebook's advertising in November). The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow.
• Source: Internal review: the Fable citation reviews of 17 October 2026 (#122) and of the Timeline, 19 October 2026 (#147), collected in #164.

• Date: 2026-10-22
• Page: Tracking & Targeting demo (/en/attention-machine/tracking-and-targeting/), its tracker test, and Privacy (/en/about/privacy/)
• What was wrong: Both said each of the test's three requests carries "this page's address". The site tells browsers to send other sites only its own address, so the requests carry https://attentionmachine.org/ and not the page's. The line shown after the test also said Google Analytics learns which page you opened and where you came from; this request tells it neither. And neither page said that, in reply, X sets a cookie called muc_ads that lasts two years and DoubleClick a test cookie that lasts fifteen minutes, if the browser accepts cookies from other sites.
• What we changed: The test's text before the button, the page's opening note and the privacy page now say "this site's address (not this page's)" and name both cookies and how long they last; the page's Sources carry the request and response headers of the run that showed it. The line after the test says Google learns that a browser at your IP address opened a page on this site, and when, and that the real tag on a site also sends which page and where you came from. The requests themselves are unchanged. All five other editions carry the same change.
• Source: Internal review (#161, found while fixing the probes in #159), against the request and response headers of a real run in Chromium on 22 October 2026.

• Date: 2026-10-22
• Page: Privacy (/en/about/privacy/)
• What was wrong: It said that list of what the site's analytics records was complete, but it left out a reading the tracker sends by itself with each page view: how far down the page was read and for how long, as a percentage and a number of seconds, with nothing that identifies the reader. It also did not say that a click on one of the links under a demo's "One thing to do now" is counted as such, or how long the counts are kept.
• What we changed: The page now says all three: the scroll depth and time on the page, the demo's links, and that the counts are kept with no end date, because Plausible as we run it has no setting that deletes them. The table row for Plausible names the reading too.
• Source: Internal review: the analytics audit of 22 October 2026 (#158, finding 3; #160).

• Date: 2026-10-22
• Page: AI & The Next Layer (/en/attention-machine/ai/)
• What was wrong: It said that since 2022 tools can produce convincing text, images, cloned voices and synthetic video "at near-zero cost, at any scale, in seconds", and further down that influence operations "can now be run at near-zero marginal cost". The only source cited, the Stanford AI Index 2023, supports the 2022 releases but says nothing about the cost, scale or speed of producing an output; the second sentence cited nothing.
• What we changed: The first sentence now says that researchers expect these tools to drive down the cost of influence campaigns and make them easier to scale; the second says such operations can now be run far more cheaply. Both link Goldstein et al. (2023), a joint report by Georgetown CSET, OpenAI and the Stanford Internet Observatory, which forecasts exactly that for language models. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts follow in the same change.
• Source: Internal review (#153), checked against the AI Index 2023 report (full text) and Goldstein et al., Generative Language Models and Automated Influence Operations (arXiv:2301.04246), p. 8.

• Date: 2026-10-22
• Page: AI & The Next Layer (/en/attention-machine/ai/)
• What was wrong: It said the EU AI Act, the DSA and national regulations were attempting to address AI-driven influence operations, and that "None are yet fully implemented or enforced." The DSA has applied in full since February 2024, and on 5 December 2025 the Commission fined X €120 million under it; Article 50 of the AI Act, its transparency duties for AI systems and AI-generated content, has applied since 2 August 2026.
• What we changed: The sentence now says the DSA is being enforced, with the X fine as the example, that Article 50 has applied since 2 August 2026, and that how far either reaches influence operations of this kind is not yet settled. Both facts link the Commission's own pages. The Dutch, Romanian and Brazilian Portuguese texts follow in the translation sync.
• Source: Internal review (#153), checked against the Commission's press release of 5 December 2025 and its FAQ on Article 50 (docs/review/cite-verify-2026-10-19/triage.md, §10).

• Date: 2026-10-22
• Page: Symptoms → Xenophobia & Violence (/en/symptoms/xenophobia/)
• What was wrong: It said both the UN Independent Fact-Finding Mission and the human-rights impact assessment Facebook commissioned from BSR "concluded that the company's response to the hate speech on its platform had been slow and inadequate". The Mission's report (A/HRC/39/64, ¶74) says "the response of Facebook has been slow and ineffective"; the BSR assessment says neither "slow" nor "inadequate" about Facebook's response. It finds that "the Facebook platform in Myanmar is being used by bad actors to spread hate speech, incite violence, and coordinate harm".
• What we changed: The sentence now quotes each source in its own words, and the Mission's entry links the report itself. The Dutch, Romanian and Brazilian Portuguese texts follow in the translation sync.
• Source: Internal review (#153), checked against the Mission's report (A/HRC/39/64, 12 September 2018) and the BSR assessment (October 2018), read in full (docs/review/cite-verify-2026-10-19/triage.md, §10).

• Date: 2026-10-22
• Page: Symptoms → Health Misinformation (/en/symptoms/health-misinformation/)
• What was wrong: It said health misinformation "also leads people to reject treatment for heart disease", citing Nielsen and Nordestgaard (2016). The study followed 674,900 people in Denmark who started statins: early discontinuation rose with negative statin-related news stories, and stopping early was associated with more heart attacks and cardiovascular deaths. It does not call the stories misinformation, and it measures stopping a treatment, not rejecting one.
• What we changed: The sentence now says what the study found, as an observed association, and introduces it as coverage that need not be false to do harm. The Dutch, Romanian and Brazilian Portuguese texts follow in the translation sync.
• Source: Internal review (#153), checked against the study's abstract (European Heart Journal 37(11), 908–916) (docs/review/cite-verify-2026-10-19/symptoms.md, H-9).

• Date: 2026-10-22
• Page: Symptoms → Consumerism (/en/symptoms/consumerism/)
• What was wrong: It said Facebook likes could predict a person's personality, politics and sexual orientation "more accurately than their friends could", citing Youyou, Kosinski and Stillwell (2015). That study compared computer and friends' judgements of personality only (r = 0.56 against 0.49); the predictions of politics and sexual orientation come from Kosinski et al. (2013), which made no comparison with friends.
• What we changed: The sentence now says "and, for personality, more accurately than their friends could". The Dutch, Romanian and Brazilian Portuguese texts follow in the translation sync.
• Source: Internal review (#153), checked against both papers' abstracts (PNAS 112(4); PNAS 110(15)) (docs/review/cite-verify-2026-10-19/symptoms.md, C-4).

• Date: 2026-10-22
• Page: Free Speech (/en/truth/free-speech/)
• What was wrong: It said that, as Susser, Roessler and Nissenbaum define it, manipulation "influences a person by exploiting weaknesses and bypassing their capacity for conscious, rational choice". Influence that circumvents rational, deliberative faculties is the non-rational-influence account their paper surveys and sets aside. Their own definition is "intentionally and covertly influencing their decision-making, by targeting and exploiting their decision-making vulnerabilities".
• What we changed: The sentence now gives their definition: manipulation is hidden influence that steers a person's decisions by targeting and exploiting their decision-making vulnerabilities, in a way they are not aware of. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts carry the same change.
• Source: Internal review (#153), checked against Susser, Roessler & Nissenbaum (2019), Technology, autonomy, and manipulation, Internet Policy Review 8(2) (docs/review/cite-verify-2026-10-19/truth-about-objections.md).

• Date: 2026-10-22
• Page: Free Speech (/en/truth/free-speech/)
• What was wrong: The glossary entry "Reverse censorship" began "Tim Wu's term for silencing speech…". Wu did not coin it: in Is the First Amendment Obsolete? (Knight First Amendment Institute, 2017) he writes of "'flooding' tactics (sometimes called 'reverse censorship')".
• What we changed: The entry now begins "A term, taken up by the legal scholar Tim Wu, for silencing speech…" and cites the essay. The Dutch, Romanian and Brazilian Portuguese texts follow in the translation sync.
• Source: Internal review (#153), checked against the essay's text on knightcolumbia.org (docs/review/cite-verify-2026-10-19/glossary-C.md).

• Date: 2026-10-22
• Page: Timeline (/en/timeline/), "A note on how this page is built" and its Sources
• What was wrong: It called the three tracks "a standard parallel-track (“swim-lane”) encoding", citing Segel and Heer (2010), and described that paper as "the basis for narrative and parallel-track timelines". The paper sets out the genres of narrative visualization; it does not discuss lanes, parallel tracks or timelines of that kind.
• What we changed: The note now says the page is a narrative timeline, one form of what Segel and Heer call narrative visualization, and that its three tracks run side by side in parallel lanes, without crediting the lanes to the paper. The entry now reads "narrative visualization and its genres". The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts carry the same change.
• Source: Internal review (#153), checked against Segel, E. & Heer, J. (2010), Narrative Visualization: Telling Stories with Data (idl.cs.washington.edu).

• Date: 2026-10-22
• Page: Timeline (/en/timeline/)
• What was wrong: The entry "Christchurch: New Zealand's Royal Commission documents the attacker's online activity" stood under 2019, the year of the attack. The Royal Commission's report was presented on 26 November 2020 and made public on 8 December 2020, and the Symptom track dates each harm to when it was first documented — in peer-reviewed research, official investigations, or court — as the Rohingya entry is dated to the UN report of 2018.
• What we changed: The entry now stands under 2020. Its wording is unchanged. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts carry the same change.
• Source: Internal review (#153); the page's own rule ("How to read this page") and the Radicalisation page's source, Royal Commission of Inquiry (2020), Ko tō tātou kāinga tēnei.

• Date: 2026-10-22
• Page: Lawmakers (/en/for-you/lawmakers/), Mental Health (/en/symptoms/mental-health/), Parents (/en/for-you/parents/)
• What was wrong: Under "what is well-established", it said "Passive social media use correlates with worse mental health outcomes in adolescents, with a large UK cohort study finding the association stronger for girls". The study (Kelly et al. 2018, UK Millennium Cohort Study, 10,904 14-year-olds) measured hours of social media use, not passive use. Whether passive use is linked to worse well-being is contested: a 2022 review of 40 survey studies found that most did not support the link. Mental Health said, under "What is well-established", that passive scrolling "correlates more strongly with negative outcomes than active social use", in the paragraph that ends "the correlation itself is not seriously disputed"; Parents said, in its opening paragraph on the research, that "other research finds it stronger for passive scrolling than active social interaction". The passive-use glossary entry on both pages called passive use "the form most consistently linked to lower well-being" without saying this is contested, and Parents listed "worse for passive use" among the established findings.
• What we changed: The sentence now says that heavier social media use correlates with worse mental health outcomes in adolescents, with the UK cohort study finding the association with depressive symptoms stronger for girls. The same sentence says that whether passive scrolling is the harmful part is contested, and cites both sides: experiments that found it lowered mood (Verduyn et al. 2015) and a review in which most studies did not find the link (Valkenburg, van Driel & Beyens 2022). On Mental Health, "not seriously disputed" now covers only the association with heavy use, and a sentence of its own says that some studies find passive scrolling linked more strongly to negative outcomes than active use, but that the finding is contested: a 2022 review of 40 survey studies found that most did not bear it out (Valkenburg, van Driel & Beyens 2022, added to the page's Sources). On Parents, the opening sentence says "some research" and marks the finding contested in the same sentence, with the same review; its Deeper box lists the passive-versus-active question among the contested points. The glossary entry on both pages now says "a 2017 review found passive use the form most consistently linked to lower well-being, though this is contested", citing Verduyn et al. 2017 and Valkenburg et al. 2022. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts carry the same changes.
• Source: Internal review (#153), checked against the abstracts of Kelly et al. 2018 (EClinicalMedicine 6), Verduyn et al. 2015 (Journal of Experimental Psychology: General 144(2)) and Valkenburg, van Driel & Beyens 2022 (New Media & Society 24(2)) (docs/review/cite-verify-2026-10-19/triage.md §10, SF-52).

• Date: 2026-10-22
• Page: Recruited to Harm (/en/symptoms/recruited-to-harm/)
• What was wrong: The glossary entry "Violence as a service" said Europol's EU-SOCTA 2025 "identified it as a fast-growing model across Europe". The assessment does not say that. It says "Organised crime-related violence is intensifying in several Member States", by "professional actors operating without borders under a violence-as-a-service model".
• What we changed: The sentence now says that EU-SOCTA 2025 found organised-crime violence intensifying in several member states under a violence-as-a-service model. The entry also cites Europol's Intelligence Notification of November 2024 for the young recruits, whose "no criminal records make them low-risk operatives". The Dutch, Romanian and Brazilian Portuguese texts follow in the translation sync.
• Source: Internal review (#153), checked against EU-SOCTA 2025 and Europol Intelligence Notification 2024-033 (docs/review/cite-verify-2026-10-19/glossary-C.md).

• Date: 2026-10-22
• Page: Tracking & Targeting (/en/attention-machine/tracking-and-targeting/), the live tracker test
• What was wrong: The test's Google Analytics and Google Ads requests used Universal Analytics, which stopped processing data on 1 July 2023. The endpoints still answered, so the blocked/loaded result was right, but the requests were not the ones Google's tags send today. The ad-exchange result also said a remarketing list is one "advertisers buy"; an advertiser builds it from its own visitors.
• What we changed: The Google Analytics request is now the one a GA4 tag sends, and the ad-exchange request is now the Google Ads remarketing tag's. The test's description now says it sends three requests rather than loading three pixels, and the ad-exchange result now says what a remarketing tag does. The Dutch, Romanian, Brazilian Portuguese, French and Spanish texts changed with it.
• Source: Internal review (#153, #159), checked against Google's Google Analytics 4 has replaced Universal Analytics, Use a Content Security Policy and Remarketing pages (docs/review/cite-verify-2026-10-19/home-faq-glossary-demos.md, G5 §8).

• Date: 2026-10-22
• Page: AI (/en/attention-machine/ai/), Elections (/en/symptoms/elections/)
• What was wrong: The Sources entries on both pages, and the glossary entry for coordinated inauthentic behaviour, labelled the Atlantic Council's analysis of the Romanian election annulment "Atlantic Council (DFRLab)". The piece (6 December 2024) is a post on the Council's New Atlanticist blog, "By Atlantic Council experts"; it is not a publication of the Digital Forensic Research Lab.
• What we changed: The entries now read "Atlantic Council (New Atlanticist)". No sentence on either page rested on the DFRLab label. The Dutch, Romanian and Brazilian Portuguese entries carry the same fix.
• Source: Internal review (#153), checked against the article's page on atlanticcouncil.org (docs/review/cite-verify-2026-10-19/glossary-A.md).

• Date: 2026-10-22
• Page: Get-Rich-Quick & Scams (/en/symptoms/get-rich-quick/), Consumerism (/en/symptoms/consumerism/)
• What was wrong: Both pages, in the glossary entry "Social proof" and in their Sources, cited Robert Cialdini's book as "Cialdini, R. (1984), Influence: The Psychology of Persuasion, Harper Business." The 1984 first edition is Influence: How and Why People Agree to Things, published by William Morrow in New York. "The Psychology of Persuasion" is the subtitle of later editions, and Harper Business published the 2021 edition.
• What we changed: The four entries now read "Cialdini, R. (1984), Influence: How and Why People Agree to Things, William Morrow." The Dutch, Romanian and Brazilian Portuguese pages carry the same fix.
• Source: Internal review (#153), checked against the Open Library record of the 1984 edition (Morrow, New York; LCCN 83021963, OCLC 10019153) and East Carolina University's catalogue record (docs/review/cite-verify-2026-10-19/glossary-C.md, Social proof; should-fix SF-76).

• Date: 2026-10-22
• Page: Elections (/en/symptoms/elections/)
• What was wrong: Its Sources entry for the BfV Verfassungsschutzbericht 2024 described the report as recording "57,701 extremism-related crimes, an all-time high". The report gives the figure and the 2023 comparison but calls no year a record; "ein Höchststand" is the federal government's wording in its account of the report's presentation (10 June 2025).
• What we changed: The entry no longer says "an all-time high". On State Actors, where the sentence about 2024 uses the words, they now link the federal government's account. The Dutch, Romanian and Brazilian Portuguese entries carry the same change.
• Source: Internal review (#153), checked against the BfV Verfassungsschutzbericht 2024 and the Bundesregierung page of 10 June 2025 (docs/review/cite-verify-2026-10-19/triage.md, §10).

• Date: 2026-10-19
• Page: Online Safety Act (/en/laws/online-safety-act/)
• What was wrong: It said Ofcom's £950,000 fine on an online suicide forum (13 May 2026) was "the first penalty for the illegal-content duties themselves". Ofcom's page says only that the forum "was the first service to be investigated"; on this page's own account, 4chan was fined £50,000 on 19 March 2026 for failing to assess the risk of illegal content. It also quoted the House Judiciary Committee's interim staff report as calling the Act "the most notable foreign attempt to imitate the Digital Services Act"; the report (3 February 2026, p. 137) says "Perhaps the most notable foreign attempt to imitate the Digital Services Act has been the United Kingdom's (UK) Online Safety Act".
• What we changed: The "first penalty" clause is removed; the sentence now gives the fine, the failures and the ten working days. The quotation now begins "perhaps the most notable foreign attempt". The Dutch, Romanian and Brazilian Portuguese texts follow in the translation sync.
• Source: Internal review (#147), checked against Ofcom's page of 13 May 2026, Ofcom's page of 19 March 2026 and the House Judiciary Committee's interim staff report (docs/review/cite-verify-2026-10-19/laws.md, must-fix 3 and 7).

• Date: 2026-10-19
• Page: Blog: "Everywhere and nowhere: what is documented about visibility before Brazil's run-off" (/en/blog/brazil-runoff-visibility-what-is-documented/)
• What was wrong: Its Sources entry attributed the Diário do Grande ABC article of 11 October 2026 (Meta alega problemas técnicos para menor alcance de Lula nas redes) to Agência Estado, and dated Meta's statement to 10 October. The article's byline is "Da Agência Brasil", and nothing in it dates the statement.
• What we changed: The entry now says Agência Brasil, as published by Diário do Grande ABC, and no longer dates Meta's statement to 10 October.
• Source: Internal review (#147), checked against the article's byline and its structured-data author (docs/review/cite-verify-2026-10-19/share-blog.md, must-fix 0(b)).

• Date: 2026-10-19
• Page: Blog: "Everywhere and nowhere: what is documented about visibility before Brazil's run-off" (/en/blog/brazil-runoff-visibility-what-is-documented/)
• What was wrong: It said the TSE minister "did not, 'at this stage', recognise any algorithmic discrimination", with "at this stage" in quotation marks as if from the order. The order has no such sentence; the words are the outlet's own summary of it.
• What we changed: The sentence now reads "He had not, at this stage, recognised any algorithmic discrimination, Congresso em Foco reported", without quotation marks. The post also now carries a dated update on the court's decision of 15 October and its confirmation on 16 October.
• Source: Internal review (#147), checked against Congresso em Foco ("não reconheceu neste momento a existência de discriminação algorítmica") and the despacho of 10 October 2026 (docs/review/cite-verify-2026-10-19/share-blog.md, must-fix 0(c)).

• Date: 2026-10-19
• Page: State Actors (/en/attention-machine/state-actors/), Elections (/en/symptoms/elections/), Lawmakers (/en/for-you/lawmakers/)
• What was wrong: It quoted the AIVD's Director-General, Simone Smit, as saying "In the eighty years that the AIVD and its predecessors have existed, there has never been a threat picture like the present one" in the AIVD's 2025 annual report. The sentence is not in the report PDF the page linked to; the foreword never says it. Elections and Lawmakers credited the same "threat picture unprecedented in 80 years" sentence to the report. State Actors also summarised the report's findings loosely, as "young people being radicalised online and deployed by state actors".
• What we changed: All three pages now credit the sentence to Smit as quoted by NOS, link the NOS article of 23 April 2026, and keep the AIVD report as the source for what the report itself documents. State Actors' sentence on the report's findings now says exactly what the report found: young people now make up a third of its jihadist caseload, and a minor mapped the digital networks of embassies and other organisations in The Hague at the request of a Russian state-backed hacking group.
• Source: Internal review (#147), checked against the NOS article of 23 April 2026 and the AIVD's 2025 annual report (aivd.nl) (docs/review/cite-verify-2026-10-19/attention-machine.md, must-fix 1).

• Date: 2026-10-19
• Page: State Actors (/en/attention-machine/state-actors/)
• What was wrong: It said one in five of the 232 terrorism arrests "in 2024" were children "under 17". The 232 arrests are the Home Office count for Great Britain in the year to March 2025, and 43 of them, 19%, were of people aged 17 or under.
• What we changed: The sentence now says "in the past year", "children" and, with a link to the Home Office statistics, 43 of 232 in Great Britain in the year to March 2025, aged 17 or under.
• Source: Internal review (#147), checked against the UK Home Office statistics on the operation of police powers under the Terrorism Act 2000, quarterly update to March 2025 (docs/review/cite-verify-2026-10-19/attention-machine.md, must-fix 5).

• Date: 2026-10-19
• Page: State Actors (/en/attention-machine/state-actors/)
• What was wrong: It said the BfV's 2024 report calls the internet a "catalyst" for radicalisation, with a particular danger for the very young. The report says that extremist chat groups serve as "catalysts" in which extreme fantasies of violence are shared, and that the internet is an easily accessible virtual space for young actors to network, in some cases even minors.
• What we changed: The sentence now says the report calls extremist chat groups "catalysts" of radicalisation and the internet an easily reached space for the very young, some of them minors.
• Source: Internal review (#147), checked against the Bundesamt für Verfassungsschutz's Verfassungsschutzbericht 2024 (docs/review/cite-verify-2026-10-19/attention-machine.md, must-fix 7).

• Date: 2026-10-19
• Page: Gamification demo (/en/attention-machine/gamification-demo/)
• What was wrong: It said, citing Susser, Roessler and Nissenbaum (2019), that the line between persuasion and manipulation is whether you are free to leave. The paper makes no such argument: it defines manipulation as influence that is hidden and that goes around a person's deliberation.
• What we changed: The sentence now says the line is whether the influence is hidden from you, whether you can see it and weigh it, and the Sources entry describes the paper the same way.
• Source: Internal review (#147), checked against Susser, Roessler and Nissenbaum, "Technology, autonomy, and manipulation", Internet Policy Review (2019) (docs/review/cite-verify-2026-10-19/attention-machine.md, must-fix 8).

• Date: 2026-10-19
• Page: AI & The Next Layer (/en/attention-machine/ai/)
• What was wrong: It called the later finding on the influencer campaign "a later tax-authority investigation". Euronews reports an investigation by the news site Snoop.ro, which cited a source at the tax authority (ANAF).
• What we changed: The page now says "a later investigation by the news site Snoop.ro, drawing on the tax authority".
• Source: Internal review (#147), checked against Euronews, 23 December 2024 (docs/review/cite-verify-2026-10-19/attention-machine.md, must-fix 9).

• Date: 2026-10-19
• Page: Timeline (/en/timeline/), the opening paragraph and its Sources
• What was wrong: It described the 2026 scrollytelling study as "a preprint, not yet peer-reviewed" and credited it to "Méndez, G.G. et al.". The paper was published at the CHI 2026 conference (13–17 April 2026), so it is peer-reviewed, and it has two authors.
• What we changed: The sentence now says "a CHI conference paper on scrolling privacy policies". The entry reads "Méndez, G.G. & Such, J. (2026)", names the proceedings of CHI '26 and the ACM DOI, and links the open version on arXiv.
• Source: Internal review (#147), checked against the arXiv record 2603.04367 ("To appear in CHI 2026") and the ACM DOI 10.1145/3772318.3790704 (docs/review/cite-verify-2026-10-19/truth-about-objections.md, must-fix 3).

• Date: 2026-10-19
• Page: Tools (/en/educate/tools/), the Ecosia entry and its Sources
• What was wrong: It said Ecosia "deletes your IP address within seven days" and passes the address on "only in shortened form, and never together with the query", with Microsoft as the destination. The wording came from a 2022 copy of Ecosia's privacy policy. The policy now in force (last modified 21 May 2026) says Ecosia anonymises the IP address after a maximum of seven days, that its search partner receives the IP address and the search terms, and that for eligible users Google results are the default, with Microsoft Bing a choice.
• What we changed: The entry says Ecosia anonymises the address within seven days, that the address and the query go to its search partner (Google by default for eligible users, Microsoft Bing if chosen), and that the partners' ads fund it. The Sources entry quotes the current policy. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), checked against Ecosia's privacy policy as last modified 21 May 2026 (docs/review/cite-verify-2026-10-19/educate-for-you.md, must-fix 9).

• Date: 2026-10-19
• Page: Lawmakers (/en/for-you/lawmakers/)
• What was wrong: It said the BfV Verfassungsschutzbericht 2024 recorded 57,701 extremism crimes, "all-time high", with "the internet" named as a "catalyst" for radicalisation. The report gives the figure and the 2023 comparison (39,433), but calls no year a record for it, and the word "Katalysatoren" is applied to online chat groups, not to the internet.
• What we changed: The sentence now says 57,701 extremism crimes, up 46% on 2023, with online chat groups named as "catalysts" for radicalisation. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), checked against the BfV Verfassungsschutzbericht 2024 (docs/review/cite-verify-2026-10-19/educate-for-you.md, must-fix 7).

• Date: 2026-10-19
• Page: Symptoms → Get-Rich-Quick & Scams (/en/symptoms/get-rich-quick/)
• What was wrong: It said the Financial Conduct Authority led "a nine-country week of action". The FCA's release says nine regulators took part, from six jurisdictions: Australia, Canada, Hong Kong, Italy, the United Arab Emirates and the United Kingdom.
• What we changed: The sentence now says "a week of action by nine regulators in six jurisdictions". The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), checked against the FCA's press release of 6 June 2025 (docs/review/cite-verify-2026-10-19/symptoms.md, must-fix 1).

• Date: 2026-10-19
• Page: Symptoms → Elections & Democratic Harm (/en/symptoms/elections/)
• What was wrong: It said a tax-authority investigation found that an influencer campaign that ended up promoting Călin Georgescu, "and whose hashtag was changed along the way", had been paid for by a rival pro-EU party. The linked Euronews report (23 December 2024) carries an investigation by the news site Snoop.ro, citing a source at the tax authority, and the payment by the PNL; it does not carry the hashtag change.
• What we changed: The sentence now credits an investigation by the news site Snoop.ro, drawing on the tax authority, and no longer mentions the hashtag. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), checked against Euronews, 23 December 2024 (docs/review/cite-verify-2026-10-19/symptoms.md, must-fix 6).

• Date: 2026-10-19
• Page: Advertising & Surveillance (/en/attention-machine/advertising/)
• What was wrong: It said Google's interest-based advertising of 2009 was "precisely the combination of online behaviour and identity profiling that advocates had warned about a decade earlier". The 2009 product profiled browsers by the sites they visited, not named people; Google kept DoubleClick's browsing records separate from the names in its users' accounts until it changed its privacy policy in 2016 (ProPublica). It also said the FTC closed its 2001 investigation "without finding a violation", which read as a clean bill: the FTC's closing letter says the closure is "not to be construed as a determination that a violation may not have occurred".
• What we changed: The 2009 sentence now says the product profiled each browser by the sites it visited, that Google kept that record apart from account names for seven years, and that the combination advocates had warned about in 1999 came in 2016. The 2001 sentence now says the FTC closed the file without taking action, while noting that closing it was no finding that the law had been kept. Both are linked to their sources, as are the dates of DoubleClick's founding, the Abacus merger and Google's purchase. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), research checked by a second model (docs/research/gaps-147-2026-10-19.md).

• Date: 2026-10-19
• Page: Journalism (/en/truth/journalism/)
• What was wrong: It called Ida Tarbell's work on Standard Oil a "two-year investigation". The series ran for two years, from November 1902 to October 1904, but Allegheny College, which holds her papers, says it came "after years of painstaking research", so the investigation was longer.
• What we changed: The sentence now says her investigation was published in nineteen parts in McClure's Magazine from 1902 to 1904, and links Allegheny College's account. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), research checked by a second model (docs/research/gaps-147-2026-10-19.md).

• Date: 2026-10-19
• Page: About (/en/about/), How We Use AI (/en/about/how-we-use-ai/), Corrections & Changes (/en/about/corrections/), Privacy (/en/about/privacy/), FAQ (/en/faq/), "Yes, but…" (/en/objections/), Translation Status (/en/about/translations/), the footer of every page, and llms.txt
• What was wrong: These pages said the project's repository on GitLab was public — that all content, the full source code and history, the rules the AI tools work under and the full list of checks could be read there — and linked into it: the footer's "Source" link, the Translation status page's issue links, and the links to AGENTS.md, REVIEW.md and the list of checks. The repository is private; a reader who followed any of these links reached a sign-in page.
• What we changed: The sentences now say what readers can see on this site: the content is published here under CC BY-NC-SA 4.0, the corrections log and the change log are published on the Corrections page, and the rules the AI tools work under are summarised on How We Use AI. The links into the repository are gone, from the footer in every language too. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), the project's visibility read from GitLab, 19 October 2026 (docs/review/cite-verify-2026-10-19/truth-about-objections.md, must-fix 1).

• Date: 2026-10-19
• Page: Timeline (/en/timeline/)
• What was wrong: "The image-first era (Instagram); targeting grows granular" was placed in 2012, which read as the year Instagram arrived. Instagram launched on 6 October 2010; what happened in 2012 is that Facebook bought it. No source on the site carried "targeting grows granular" as a 2012 event.
• What we changed: The 2012 entry now reads "Facebook buys Instagram, the photo-sharing app built for the phone". The Consumerism page, where it points, now says so and cites Facebook's announcement of 9 April 2012 and its Form 10-Q for the third quarter of 2012. The 2007 entry no longer says the iPhone "puts the feed in every pocket" (in 2007 it went on sale in the US only); it quotes Apple's own words, "the Internet in your pocket", sourced on the same page. The Dutch, Romanian and Brazilian Portuguese drafts carry the same changes.
• Source: Internal review (#147), research checked by a second model (docs/research/timeline-147-2026-10-19.md).

• Date: 2026-10-19
• Page: Get Help (/en/educate/get-help/) and every page that lists its entries
• What was wrong: It listed "Action Fraud — the UK's national fraud reporting centre", which Report Fraud replaced on 4 December 2025; the new service covers England, Wales and Northern Ireland only, so "the UK's" was wrong too. It said the Saudi Ministry of Health's 937 line gives "medical or psychological advice"; the ministry's own pages describe medical consultations only. The link for India's 112 pointed at a page that no longer exists.
• What we changed: The entry is now "Report Fraud, formerly Action Fraud", for England, Wales and Northern Ireland, with Scotland reporting via 101, and links to reportfraud.police.uk. The 937 entry says "medical advice". The 112 entry links to the ministry's current page. The Dutch, Romanian and Brazilian Portuguese texts carry the same changes.
• Source: Internal review (#147), checked against gov.uk's announcement of 4 December 2025, Report Fraud's own site (as archived on 11 October 2026) and the Ministry of Health's contact, 937 and psychiatry pages (docs/review/cite-verify-2026-10-19/educate-for-you.md, must-fix 2, 4 and 5).
• Update (2026-10-22): The 937 entry says "medical or psychological advice" again. The National Center for Mental Health Promotion, a government body, says on its own FAQ that psychological consultations are requested through 937 (https://ncmh.org.sa/faqs, re-read on 22 October 2026). The Ministry of Health's own pages still describe medical consultations only.

• Date: 2026-10-19
• Page: See it work (/en/demos/) and Tracking & Targeting demo (/en/attention-machine/tracking-and-targeting/)
• What was wrong: Both opened by saying nothing is sent anywhere. The tracker test in the first demo loads three real tracking pixels, from Google Analytics, Google Ads / DoubleClick and X Ads, when the reader presses its button, and each request carries the reader's IP address, the page's address and any cookie the browser already holds for that company. The test's own text said so; the pages above it did not.
• What we changed: The lead of both pages and the description of the first now say that nothing leaves the browser except the tracker test, which sends three real requests to Google and X when its button is pressed. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), against the test's own copy and its list of hosts (docs/review/cite-verify-2026-10-19/home-faq-glossary-demos.md, must-fix 2).

• Date: 2026-10-19
• Page: Tracking & Targeting demo (/en/attention-machine/tracking-and-targeting/), the line shown after the live test
• What was wrong: After the test the line for Google Analytics said it learns "which other sites you've visited", "tied to a cross-site ID". The page's text had been corrected on 17 October (C-307): Google's cookie is set separately on each site, but on every one of them the browser sends Google the page's address and the reader's IP address.
• What we changed: The line now says Google learns that you opened the page, when and from what referrer, plus your IP address and the page's address, which Google may use to personalise the ads you see elsewhere. All five other editions' strings carry the same change.
• Source: Internal review (#147), against Google Analytics Help, [GA4] Cookie usage on websites (docs/review/cite-verify-2026-10-19/home-faq-glossary-demos.md, must-fix 1).

• Date: 2026-10-19
• Page: Donate (/en/share/donate/), Sources
• What was wrong: The Ko-fi entry said the platform charges no fee on one-off donations. Ko-fi's pricing page says its free plan charges no service fee on tips and its Standard plan, the default for new accounts, charges 5% on all payment types including tips; the payment processor's fee applies on either.
• What we changed: The entry now gives both plans and the processor's fee. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#147), against Ko-fi's pricing page as archived on 13 October 2026 (docs/review/cite-verify-2026-10-19/share-blog.md, must-fix 2).

• Date: 2026-10-17
• Page: Symptoms → Body Image (/en/symptoms/body-image/)
• What was wrong: It said maxillofacial surgeons "have warned repeatedly" that bonesmashing risks "broken bones, nerve and dental injury, lasting asymmetry, and disfigurement", with no source for the sentence. The warnings by oral and maxillofacial surgeons that could be found name broken facial bones, nerve damage, dental injury, bite changes and asymmetric healing; "permanent disfigurement" was a plastic surgeon's word, not theirs.
• What we changed: The sentence now says they have warned "in their specialty's journal and in the press", lists only the harms those surgeons name, and links two reports quoting them (VICE, 1 March 2026; CBS 8, 27 March 2026). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Later change: the unsourced "surgeons describe as reckless" removed from the sentence before it, in all four languages.
• Source: Internal review (#122, step 3), checked against the two reports and the specialty journal's record on PubMed (docs/review/source-links-step3/open-phrases.md, verified).

• Date: 2026-10-17
• Page: Symptoms → Mental Health (/en/symptoms/mental-health/)
• What was wrong: It said Candice Odgers' review of The Anxious Generation in Nature argues "that most of the causal experiments it cites were run on adults". The review (Nature 628, 29–30, 2024) argues that most of the data are correlational and that an analysis across 72 countries finds no consistent link with the roll-out of social media; it says nothing about experiments on adults.
• What we changed: The clause is removed; the sentence keeps the timing argument and the other causes of the decline. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change. Yes, but… carried the same clause and is corrected in its own section's change.
• Source: Internal review (#122, step 3), checked against the review's full text (the same check as for Yes, but…).

• Date: 2026-10-17
• Page: Symptoms → Body Image
• What was wrong: After noting that Dove's survey is brand-funded, it said the direction of Dove's findings "matches independent work, such as that of the Centre for Appearance Research at the University of the West of England, which studies the same effect with no product to sell". The Centre has co-created the Dove Self-Esteem Project's body-confidence programmes with Dove since 2014 (its own page on the partnership), and the entry linked only the Centre's front door, which carries no finding.
• What we changed: The page now says that the Centre has worked with Dove since 2014, and rests the independent support on a meta-analysis of 63 samples (Saiphoo & Vahedi, 2019: a small but significant link between social-media use and body-image disturbance). The Centre's entry links its page on the partnership. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Symptoms section (#122, step 3), checked against the Centre's page on the Dove Self-Esteem Project.

• Date: 2026-10-17
• Page: Symptoms → Consumerism
• What was wrong: It gave "10% of global emissions" and "second-largest consumer of water" as the UN Environment Programme's 2018 figures. UNEP's page of 12 November 2018 said fashion produces "10 per cent of global carbon emissions" and that "textile dyeing is the second largest polluter of water globally" — not that fashion is the second-largest consumer of water. The same UNEP page now says "between 2 to 8 per cent". The claim that the figures "have since been widely challenged" had no source on the page.
• What we changed: The sentence quotes UNEP's own words, says the page now gives 2 to 8 per cent, and links the challenge to Alden Wicker's Vox investigation (2020). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the UNEP page as published (Wayback Machine capture of 12 November 2018) and as it stands, and against the Vox article.

• Date: 2026-10-17
• Page: Symptoms → Consumerism
• What was wrong: "Research repeatedly finds associations between heavy social media use and compulsive buying, higher levels of materialism, and reduced satisfaction with owned possessions", and the Deeper box repeated the last item. The one study cited (Sharif & Khanekharab, 2017, a survey of 501 young adults) finds compulsive buying and materialism; it says nothing about satisfaction with possessions, and one study is not "repeatedly".
• What we changed: "Sharif and Khanekharab's 2017 survey of 501 young adults found associations between heavy social media use, compulsive buying and higher levels of materialism" — one study, named as one; the Deeper box drops the satisfaction claim. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the study's abstract.

• Date: 2026-10-17
• Page: Symptoms → Elections & Democratic Harm; Symptoms → Xenophobia & Violence
• What was wrong: Both said a 2016 internal Facebook study found "64% of the people who joined extremist groups had been led there by the platform's own recommendation tools". The slide, as the Wall Street Journal quotes it, says "64% of all extremist group joins are due to our recommendation tools" — a share of joins, not of people (one person can join several groups).
• What we changed: Both pages now say "64% of all joins to extremist groups came through the platform's own recommendation tools", as How It Works has since the same day. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the Wall Street Journal article of 26 May 2020.

• Date: 2026-10-17
• Page: Symptoms → Elections & Democratic Harm
• What was wrong: The sentence on the reach of Russia's Internet Research Agency rested on one report of the 31 October 2017 hearing (PBS NewsHour / AP), and the entry said it carried "about 20 million on Instagram". It does not: Facebook's general counsel gave the Instagram figure to the Senate Intelligence Committee the next day, and Twitter's 1.4 million notifications came in January 2018, after the testimony.
• What we changed: "About 20 million on Instagram" links Reuters' report of the 1 November 2017 hearing; "Twitter later notified about 1.4 million people" links Twitter's own update (19 January 2018, updated 31 January 2018); the PBS entry lists what it carries (126 million on Facebook; Google's 18 YouTube channels, 1,108 videos, 309,000 US views). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the PBS/AP report, Reuters (via Fortune) and Twitter's blog post (Wayback Machine capture).

• Date: 2026-10-17
• Page: Symptoms → The Manosphere
• What was wrong: It said Kevin Roose's "The Making of a YouTube Radical" (New York Times, 2019) followed one person's history "from mainstream political content to far-right extremism over eighteen months", here and in the Deeper box. The article has no eighteen months: it starts from a self-help video YouTube recommended to Caleb Cain in late 2014, follows him into far-right politics through 2015 and 2016, and rests on his history of more than 12,000 videos going back to 2015.
• What we changed: Both places now describe that span. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change. The Watch & Listen page still says "eighteen months"; it is corrected in its own section's change.
• Source: Internal review (#122, step 3), checked against the article (Wayback Machine capture of the New York Times page).

• Date: 2026-10-17
• Page: Symptoms → Mental Health
• What was wrong: It said the decline in adolescent mental health from around 2012 "corresponds with the period when smartphone ownership among teenagers crossed 50%". In Pew Research Center's surveys, 37% of US teens had a smartphone in 2012 and 73% had access to one by early 2015; the 50% mark was crossed between them, not in 2012.
• What we changed: The sentence now gives Pew's two figures, linked to the two reports. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against Pew's Teens and Technology 2013 and Teens, Social Media & Technology Overview 2015.

• Date: 2026-10-17
• Page: Symptoms → Xenophobia & Violence
• What was wrong: It said "Facebook launched in Myanmar in 2012" and became the information platform "for most of the country's population", and that "the UN" described the 2017 campaign against the Rohingya as bearing "the hallmarks of genocide". Facebook had no launch in Myanmar — its dominance took root "not by design" as the country opened up — and in 2018 its 18 million users were about a third of the population (Reuters); for most of the people online, Facebook was the internet (BSR). The phrase was said on 12 March 2018 by the UN Special Rapporteur on Myanmar, Yanghee Lee, an independent expert, not by the UN as a body; the UN Fact-Finding Mission later called for senior military officials to be investigated and prosecuted for genocide.
• What we changed: The paragraph now says this, linked to Reuters' special report (15 August 2018), the BSR assessment and the Special Rapporteur's statement. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the Reuters report, the BSR assessment, the OHCHR press release of 12 March 2018 and the Fact-Finding Mission's report (A/HRC/39/64).

• Date: 2026-10-17
• Page: Symptoms → Recruited to Harm
• What was wrong: The entry for Europol's "Operational Taskforce GRIMM: 193 arrests in 6 months" was dated 2026. Europol published it on 8 December 2025. The page also said the reports were "backed by investigations across dozens of countries", a figure no source on the page gave.
• What we changed: The entry is dated 8 December 2025. The sentence now says "a Europol-coordinated effort involving 28 countries", linked to Europol's TE-SAT 2026 (Project Compass: "law enforcement authorities from 28 Member States and partner countries"). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against Europol's own pages and the TE-SAT 2026 report.

• Date: 2026-10-17
• Page: Laws & Regulation → Courts & Liability
• What was wrong: It said that in October 2022 "hundreds of suits by families and school districts" were pooled into MDL 3047 — "the largest coordinated legal action against social media platforms in history". The Judicial Panel's order of 6 October 2022 centralised 28 actions in 17 districts, with 56 related actions notified since the motion; the school districts' suits were transferred in later. No source on the page, or found, states the superlative. The source entry also linked a December 2023 order on one school district, not the order that created MDL 3047.
• What we changed: "Dozens of personal-injury suits over young people's use of social media were pooled into multidistrict litigation (MDL 3047)… Suits by school districts were added later, and by September 2026 3,383 lawsuits had been filed in it." The entry now links the order of 6 October 2022; the December 2023 order is a separate entry for the school districts. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Laws section (#122, step 3), checked against the Judicial Panel's orders and its list of pending dockets of 1 September 2026.

• Date: 2026-10-17
• Page: Laws & Regulation → Courts & Liability
• What was wrong: It described the 51 attorneys general who settled with Meta on 26 August 2026 as "the states that had sued, joined by the rest, the District of Columbia and the territories". The settlement agreement lists 47 states, the District of Columbia, Puerto Rico, American Samoa and the Northern Mariana Islands: three states and two territories are not in it.
• What we changed: "51 US attorneys general (47 states, the District of Columbia, Puerto Rico, American Samoa and the Northern Mariana Islands)". The source entry now says the same, and explains that the 29 states in Hunton's title are those whose federal case was on trial. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Laws section (#122, step 3), checked against the settlement agreement (its definition of the Settling States' attorneys general) and the New Jersey Attorney General's release.

• Date: 2026-10-17
• Page: Laws & Regulation → Courts & Liability
• What was wrong: It put "malice, fraud or oppression" in quotation marks as the jury's finding; no source carries those words in that form — the Associated Press gives the finding, unquoted, as "malice, oppression or fraud". It also said, in the site's own voice, that the plaintiff "became addicted to Instagram and YouTube as a child"; that was her testimony. And it said Instagram was found 70% responsible; the jury apportioned it to Meta.
• What we changed: "a Los Angeles jury awarded $6 million to a plaintiff who testified that she had become addicted to social media as a child, finding that both companies acted with malice, oppression or fraud. Meta, which owns Instagram, was found 70% responsible; YouTube 30%." The verdict now has its own source, the Associated Press report of 25 March 2026. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Laws section (#122, step 3).

• Date: 2026-10-17
• Page: Laws & Regulation → Courts & Liability; Section 230
• What was wrong: Courts said "the causal pathway from social media design to psychological harm is real but more complex"; Section 230 said recommended content "contributes to radicalisation, depression, or harm". Both stated as fact a causal link the site holds to be contested (Mental Health; Radicalisation Pipelines).
• What we changed: Courts: "Whether social media design causes psychological harm, and how far, is contested: the link is well documented as a correlation, the pathway is more complex, and individual variation is large." Section 230: "content that plaintiffs and some studies link to radicalisation, depression, or harm, though whether it causes them is contested." The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Laws section (#122, step 3).

• Date: 2026-10-17
• Page: Laws & Regulation → Digital Services Act
• What was wrong: It said US Vice President Vance criticised the DSA and the GDPR as over-regulatory at the Paris AI Action Summit and the Munich Security Conference in 2025. That criticism was made in Paris, on 11 February 2025; the Munich speech of 14 February names neither law — it was about censorship and free speech.
• What we changed: "At the Paris AI Action Summit in February 2025, US Vice President Vance criticised the DSA and the GDPR as over-regulatory", sourced to the remarks as delivered (The American Presidency Project). The same paragraph now gives Meta's 2025 revenue ($201 billion, from its own filing) and 6% of it (about $12 billion) in place of an unsourced "potentially exceeding €9 billion", which was 6% of the year before, converted at an unstated rate. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Laws section (#122, step 3), checked against the transcripts of both speeches.

• Date: 2026-10-17
• Page: Laws & Regulation → Age Restrictions
• What was wrong: It said Utah in 2023 became "the first US state to require age verification for under-18s". Utah's S.B. 152 requires the age of anyone opening a social media account to be verified, and a parent's consent for under-18s — and Utah repealed it in 2024. It partly replaced it with the Utah Minor Protection in Social Media Act, which a federal court blocked in September 2024, before it took effect; Utah has appealed. The page said none of this, so it read as if the 2023 requirement were in force. It also called Australia's law "the most comprehensive national age restriction law in the world", which no source on the page states.
• What we changed: "Utah passes the first US state law requiring age verification for anyone opening a social media account, and a parent's consent for under-18s", sourced to the enrolled bill and NPR's report of 24 March 2023, followed by: "In 2024 Utah repealed it and partly replaced it with a new act, which a federal court blocked in September 2024, before it took effect; Utah has appealed", sourced to the District of Utah's order of 10 September 2024 (NetChoice v. Reyes) and Courthouse News Service's report of the appeal hearing (20 November 2025). Australia's law is "the first national law of its kind", as the OECD's account of it supports. The date the Australian rule took effect (10 December 2025) now has its own source. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Laws section (#122, step 3).

• Date: 2026-10-17
• Page: For You → Lawmakers
• What was wrong: It said, citing MI5 (October 2025), that "one in five of 232 terrorism arrests were children under 17; online radicalisation identified as a central factor", and the source entry put the arrests in 2024. The 232 arrests are the Home Office's count for the year to 31 March 2025, in Great Britain, and the age band is 17 and under (43 of 232, 19%); in calendar 2024 the share was 16%. MI5's Director General does not call online radicalisation "a central factor"; he says terrorism "breeds in squalid corners of the internet".
• What we changed: The sentence now quotes MI5 and gives the Home Office figure with its period and place, as Yes, but… does since the same day; the Home Office statistics are a new Sources entry, and the MI5 entry quotes the speech of 16 October 2025. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), following the same correction on Yes, but…, checked there against the Home Office statistics of 12 June 2025 and 13 March 2025 and the MI5 speech.

• Date: 2026-10-17
• Page: Educate → Tools & Sites; For You → Educators, Parents, Teens
• What was wrong: Four pages said the Bad News game takes ten minutes ("A 10-minute browser game", "10 minutes, browser-based", "in ten minutes", "for ten minutes"). The study of the game the site cites, Roozenbeek & van der Linden (2019), says it "takes approximately 15 minutes to complete".
• What we changed: Each page now says about 15 minutes, linked to that study where the page lists it; the study's entry notes the playing time. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change. The same figure on Tracking & Targeting, Yes, but… and Share is corrected with those sections.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against the paper (doi.org/10.1057/s41599-019-0279-9).

• Date: 2026-10-17
• Page: For You → Educators
• What was wrong: It said that resistance to a school phone policy "tends to come from parents rather than students", and that framing a ban as a wellbeing measure "produces better stakeholder engagement". Neither sentence had a source. New Zealand's national evaluation of its 2024 ban (Education Review Office, 2025) found that only 37% of Year 12–13 students follow the rules, that students are more likely to break them where parents resist, and that compliance is higher where students understand the purpose of the rules; it found nothing on wellbeing versus discipline framing.
• What we changed: The two sentences are replaced by what that review found, with the review in the Sources. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against the Education Review Office summary (Do not disturb, 2025).

• Date: 2026-10-17
• Page: For You → Educators
• What was wrong: It said the Bad News game is "recommended for ages 15 and up" and is "used in classrooms in several countries". The game's makers say on its own site, and in their info sheet for educators, that it was written for ages 14 and up; no source was found for the classroom count. (A correction of September 2026 had changed "from 14" to 15 and up; the makers' current pages say 14.)
• What we changed: The page now says "written for ages 14 and up, its makers say, with an info sheet for using it in class", linked to the game's site, and the Junior version for ages 8–11 is linked to the European Digital Media Observatory's page on it. The unsourced classroom sentence is gone. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against getbadnews.com and its educators' info sheet.

• Date: 2026-10-17
• Page: For You → Parents
• What was wrong: It said that "monitoring apps and surveillance software damage trust without producing the outcomes parents hope for" — a causal claim with no source. The two studies on it (Ghosh et al., CHI 2018) found that children's reviews of such apps said they hurt their relationships with their parents, and, in a survey of 215 parent–teen pairs, that more parental control went with more online risk, not less — an association the authors note may run from risk to monitoring.
• What we changed: The sentence now says what the two studies found, with both in the Sources and the caveat on the direction in the same sentence. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against the two papers' abstracts.

• Date: 2026-10-17
• Page: For You → Teens
• What was wrong: It said that "playing the role of the manipulator is more effective than reading about manipulation". In the research this is a hypothesis the Bad News paper quotes from McGuire ("may be more effective"); the one head-to-head comparison found, Basol et al. (2021), saw a similar game do better than reading infographics without testing the difference.
• What we changed: The page now says that playing the manipulator "seems to work better", names the one study and says the gap was not tested, with that study in the Sources. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against Roozenbeek & van der Linden (2019) and Basol et al. (2021).

• Date: 2026-10-17
• Page: Educate → Watch & Listen
• What was wrong: It described Kevin Roose's "The Making of a YouTube Radical" (New York Times, 2019) as one person's YouTube history "followed over eighteen months", from mainstream content to far-right extremism "through the platform's own recommendations", and as "documenting the recommendation pipeline". The article has no eighteen months: it starts from a self-help video YouTube recommended to Caleb Cain "in late 2014", follows him into far-right politics through the following years, draws on "a log of more than 12,000 videos and more than 2,500 search queries dating to 2015", and says he swore off the alt-right "nearly five years after discovering it". (The Manosphere page carried the same error and is corrected in its own section's change.) Whether the recommendations drive such a path is contested.
• What we changed: The entry now gives the span the article gives — the same wording as on The Manosphere — calls the piece a single case, and says in the same sentence that whether the recommendation engine drives such escalation on its own is contested. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against the article (Wayback Machine copy of 8 June 2019).

• Date: 2026-10-17
• Page: Educate → Books
• What was wrong: It said Merchants of Doubt (Oreskes & Conway, 2010) traces manufactured doubt "from tobacco to climate to vaccines". The book's chapters cover tobacco and secondhand smoke, the Strategic Defense Initiative, acid rain, the ozone hole, global warming and DDT; vaccines are not among them.
• What we changed: It now says "from tobacco to climate change; the same playbook recurs around vaccines", which keeps the site's own observation apart from the book's subject. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against the publisher-listed contents (Publishers Weekly review).

• Date: 2026-10-17
• Page: For You → Lawmakers
• What was wrong: It said the CJEU "upheld key parts" of the Belgian Data Protection Authority's 2022 decision on IAB Europe's consent framework, and that its ruling "found the framework … raises serious GDPR problems". Case C-604/22 (7 March 2024) is a preliminary ruling on questions referred by the Brussels Court of Appeal, which was hearing IAB Europe's challenge; the Court of Justice "does not decide the dispute itself". It held that the consent string is personal data and that IAB Europe can be a joint controller, subject to the referring court's checks.
• What we changed: The page now says the CJEU, on questions referred by a Belgian court, confirmed key parts of that reasoning, and the definition says the ruling held that the consent string is personal data. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against CJEU press release No 44/24 (7 March 2024).

• Date: 2026-10-17
• Page: Educate → What to Do; For You → Parents
• What was wrong: Both said sleep is "the best-documented pathway" from heavy social media use to worse mental health (on Parents: to anxiety and depression), stating as settled the causal link the site marks contested elsewhere. The meta-analysis they cite (Carter et al., 2016) is about sleep, not about whether use causes the harm.
• What we changed: Both now call sleep the best-documented of the pathways proposed, and say in the same sentence that the causal link is contested. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3) (the site's rule that a contested claim is marked in the same sentence).

• Date: 2026-10-17
• Page: For You → Young Adults
• What was wrong: It said "the Cambridge field experiments found it works across the political spectrum, and that it holds up over time". The field experiments (Roozenbeek et al., 2022) carry the first half; the durability finding is Maertens et al. (2021), three longitudinal experiments, which found the effect stable for at least three months with regular testing and fading within two months without. The definition of inoculation said the knowledge simply "lasts".
• What we changed: The sentence names each study's finding separately, and both it and the definition now say the effect lasts for months when refreshed; the definition's own Sources list both papers. The page also no longer uses "the manosphere pipeline" as a shorthand. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Educate and For You sections (#122, step 3), checked against both papers (Europe PMC).

• Date: 2026-10-17
• Page: Yes, but… (/en/objections/), "Children today are digital natives"
• What was wrong: It said MI5 reports that "1 in 5 UK terrorism arrests in 2024 were of children under 17, with online radicalisation a central factor". The 232 arrests behind MI5's "one in five" are the Home Office's count for the year to 31 March 2025, in Great Britain, and the age band is 17 and under (43 of 232, 19%). In calendar 2024 the share was 16%. MI5's Director General does not call online radicalisation "a central factor"; he says terrorism "breeds in squalid corners of the internet".
• What we changed: "Home Office figures show that 43 of the 232 terrorism arrests in Great Britain in the year to March 2025 — nearly one in five — were of children aged 17 or under; MI5 says terrorism now 'breeds in squalid corners of the internet'." The Home Office statistics are a new Sources entry, and the MI5 entry quotes the speech of 16 October 2025. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Truth, Timeline, About and Objections pages (#122, step 3), checked against the Home Office statistics of 12 June 2025 and 13 March 2025 and the MI5 speech.

• Date: 2026-10-17
• Page: Yes, but… (/en/objections/), "Haidt has been debunked"
• What was wrong: It said Candice Odgers' review in Nature argues "that most of the causal experiments the book cites were run on adults". The review (Nature 628, 29–30, 2024) argues that most of the data are correlational and that an analysis across 72 countries finds no consistent link with the roll-out of social media; it says nothing about experiments on adults.
• What we changed: The clause is removed; the sentence keeps the timing argument, which the review does make. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the review's full text.

• Date: 2026-10-17
• Page: Yes, but… (/en/objections/), "What can I actually do about it?"
• What was wrong: It said the Bad News game teaches six manipulation techniques "in ten minutes". Its evaluation (Roozenbeek & van der Linden 2019) describes a game "that takes approximately 15 minutes to complete".
• What we changed: "in about 15 minutes"; the Sources entry says so too. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the 2019 paper.

• Date: 2026-10-17
• Page: Timeline (/en/timeline/)
• What was wrong: "Google AdWords turns attention into a real-time auction" was placed in 2000. AdWords launched in October 2000, selling keyword ads with no auction; Google began selling its search ads by per-click auction with AdWords Select in February 2002, and Overture had done so since 1997.
• What we changed: The 2000 entry now reads "Google AdWords launches: ads sold against the words people search for", and a new 2002 entry reads "Google starts selling its search ads by auction, priced per click". The Consumerism page, where both entries point, now says the same and cites Google's two announcements and Edelman, Ostrovsky & Schwarz (2007). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against Google's press releases of 23 October 2000 and 20 February 2002 and the American Economic Review paper.

• Date: 2026-10-17
• Page: Timeline (/en/timeline/)
• What was wrong: "UN investigators link Facebook to violence against the Rohingya in Myanmar" was placed in 2017, the year of the military's campaign. The UN Independent Fact-Finding Mission's report that made the link (A/HRC/39/64) is dated 12 September 2018.
• What we changed: The entry moved to 2018. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the report.

• Date: 2026-10-17
• Page: What is Truth? → Journalism
• What was wrong: It said the Panama Papers reporting had returned $1.36 billion to public treasuries, in the lead and in the Panama Papers paragraph. That was ICIJ's tally of April 2021. ICIJ's tally of April 2025 is $1.3 billion, lower because governments corrected their own figures, and ICIJ says it almost certainly understates the total. The Sources entry linked only the investigation's landing page, which gives no figure.
• What we changed: Both sentences give $1.3 billion and attribute it to ICIJ's count (April 2025); the entry now also cites ICIJ's article of 3 April 2025. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against ICIJ's articles of 6 April 2021 and 3 April 2025.

• Date: 2026-10-17
• Page: What is Truth? → Democracy
• What was wrong: It put "the worst form of government except for all the others" in quotation marks as Churchill's line. That is the common compression; Hansard (11 November 1947, col. 207) records "democracy is the worst form of Government except all those other forms that have been tried from time to time".
• What we changed: The quotation now gives Hansard's words. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against Hansard.

• Date: 2026-10-17
• Page: About → How We Communicate
• What was wrong: "Cognitive load theory (Sweller et al.)" linked to Sweller (1988), which has one author.
• What we changed: "Sweller". The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review (#122, step 3), checked against the paper's Crossref record.

• Date: 2026-10-17
• Page: Share & Support → Share This
• What was wrong: It said the Bad News game "is 10 minutes", and its Deeper box called it "a ten-minute game". The study that evaluated the game describes it as one "that takes approximately 15 minutes to complete".
• What we changed: "It takes about 15 minutes"; "a fifteen-minute game". The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Share section (#122, step 3), checked against Roozenbeek & van der Linden (2019), Palgrave Communications 5(65).

• Date: 2026-10-17
• Page: Share & Support → Organisations Worth Supporting
• What was wrong: Its Deeper box said that correcting individual false claims "sometimes entrenches" the falsehood, with no source and no sign that this is disputed. The effect was reported by Nyhan and Reifler (2010); a later set of five experiments by Wood and Porter (2019), with more than 10,100 people and 52 issues, found no correction that triggered it.
• What we changed: The sentence now says corrections have entrenched the falsehood in some experiments, names the "backfire effect" as Nyhan and Reifler's, says it is contested, and gives Wood and Porter's result; both papers are in the page's Sources. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Share section (#122, step 3), checked against both papers.

• Date: 2026-10-17
• Page: Share & Support → Organisations Worth Supporting
• What was wrong: The Sources entry for Roozenbeek, van der Linden et al. (2022), Science Advances, described it as "the Bad News research". The paper tests five short prebunking videos, in the lab and in a field study on YouTube; it is not an evaluation of the Bad News game.
• What we changed: The entry now describes the video study; the page's mention of "the Bad News evaluations" links to the study that evaluated the game, Roozenbeek & van der Linden (2019). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Share section (#122, step 3), checked against the paper's full text.

• Date: 2026-10-17
• Page: The Attention Machine → Tracking & Targeting
• What was wrong: It said the Bad News game "teaches it interactively in ten minutes". The peer-reviewed evaluation of the game describes "a freely accessible browser game that takes approximately 15 minutes to complete".
• What we changed: "about fifteen minutes", linked to Roozenbeek & van der Linden (2019), Palgrave Communications 5(65), now in the page's Sources. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3), checked against the paper.

• Date: 2026-10-17
• Page: The Attention Machine → AI & The Next Layer
• What was wrong: It gave Farooq et al. (2025), the study of citizens' concerns about AI-generated content in the 2024 European elections, as published in Political Communication. Its DOI is in The International Journal of Press/Politics (2 October 2025). The "testing ground" description of those elections, which the page did not attribute, is the first line of the same study's abstract.
• What we changed: The sentence and the source entry name The International Journal of Press/Politics, and "testing ground" now links to that study. CIGI's Then and Now (August 2025), which uses "testing ground" only of local campaigns, stays in the Sources as further reading. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3), checked against the DOI's publisher record.

• Date: 2026-10-17
• Page: The Attention Machine → How It Works
• What was wrong: It said the 2016 Facebook presentation found "64% of the people who joined extremist groups on the platform had been led there" by Facebook's recommendation tools. The slide, as the Wall Street Journal quotes it, says "64% of all extremist group joins are due to our recommendation tools": a share of joins, not of people — one person can join several groups. The page's own Deeper box already said joins.
• What we changed: "64% of all joins to extremist groups on the platform came through Facebook's own recommendation tools." The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3), checked against the Wall Street Journal article of 26 May 2020.

• Date: 2026-10-17
• Page: The Attention Machine → Tracking & Targeting; Advertising & Surveillance
• What was wrong: Both said that in 2024 the Court of Justice of the European Union "upheld key parts" of the Belgian Data Protection Authority's 2022 decision (Advertising: "of that reasoning"). The Court did not hear an appeal: in Case C-604/22 it answered questions referred by the Brussels Court of Appeal, which was hearing IAB Europe's challenge, and the dispute itself stayed with the Belgian court.
• What we changed: Tracking & Targeting now says the Court, answering questions from the Belgian appeal court, confirmed the decision's two premises: the consent string is personal data, and IAB Europe is, subject to that court's checks, a joint controller of it. Advertising says the Court, on questions referred by a Belgian court, "confirmed" key parts of that reasoning. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3), checked against the Court's press release No 44/24 (7 March 2024).

• Date: 2026-10-17
• Page: The Attention Machine → Your Brain
• What was wrong: It described the roughly 105,000 headline variants in Robertson et al. (2023) as tests that "news publishers themselves" ran. They all come from one news site, Upworthy, whose archive of A/B tests the study analysed.
• What we changed: "about 105,000 headline variants that one news site, Upworthy, A/B-tested on its readers"; the source entry names Upworthy's archive. The numbers were right and are unchanged. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3), checked against the paper in Nature Human Behaviour.

• Date: 2026-10-17
• Page: The Attention Machine → Search
• What was wrong: It said Google's AI Overviews "have repeatedly been found to contain errors, to mix reliable and unreliable sources, and to present contested claims as settled". The one source on the page for it, Google's own account of the May 2024 launch, documents errors and answers drawn from troll posts; it does not show a repeated pattern or contested claims presented as settled, and the page cited nothing that did.
• What we changed: The page now says what Google acknowledged in May 2024: that some AI Overviews misread web pages and presented inaccurate information, and that others drew on sarcastic or troll-y forum posts, among them advice to use glue to make cheese stick to pizza. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3).

• Date: 2026-10-17
• Page: The Attention Machine → Tracking & Targeting
• What was wrong: It said Google Analytics' identifier "is shared across" the sites that use it. Google Analytics sets a first-party cookie, separately on each site. What makes it cross-site is that every page carrying it has your browser send Google the page's address and your IP address.
• What we changed: The page now says that: the cookie is set separately on each site, but on every one of them your browser sends Google the address of the page and your IP address, which Google may use, depending on your settings, to personalise ads elsewhere — the raw material of a cross-site record. Both statements are sourced to Google's own documentation. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3).

• Date: 2026-10-17
• Page: The Attention Machine → How It Works
• What was wrong: It said the feed's loop "runs continuously, invisibly, across three billion users simultaneously", with no source. The nearest documented figure is Meta's own count of people using its apps on an average day — not concurrent users, and not all platforms.
• What we changed: "for more than three billion people a day on Meta's apps alone", sourced to Meta's results for the second quarter of 2026 (3.60 billion daily active people in June 2026, as filed with the US SEC). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking review of the Attention Machine section (#122, step 3).

• Date: 2026-10-15
• Page: Courts (/en/laws/courts/)
• What was wrong: It said more than 40 US states sued Meta "in the same court" in October 2023.
• What we changed: 33 states sued together in the federal court in Northern California; eight more filed in their own state courts. The page now cites the California Attorney General's announcement.
• Source: internal review (Fable read of the Courts chronology)

• Date: 2026-10-15
• Page: Laws & Regulation → GDPR
• What was wrong: The list of fines said "2021: Luxembourg fines Amazon €746 million" with nothing after it. On 12 March 2026 Luxembourg's Administrative Court annulled that decision in full: it confirmed the breaches of the GDPR, but sent the case back to the data protection authority (CNPD), which had not examined intent or negligence or weighed measures other than a fine.
• What we changed: The list now says the decision was annulled on appeal in March 2026, that the court confirmed the breaches, and that it sent the case back to the regulator, sourced to the court's own summary (Cour administrative, arrêt n° 52757C du rôle). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: New research — the GDPR fines research note of 15 October 2026 and its independent verification (docs/research/gdpr-fines-vs-turnover-2026-10-15.md, …-verify.md).

• Date: 2026-10-15
• Page: Laws & Regulation → GDPR
• What was wrong: "Fines, large in absolute terms, are manageable as a business cost for the largest platforms", and in the Deeper box, that a record fine "functions more like a cost of doing business than a deterrent". Neither was sourced, and both stated as fact what the research treats as an open question: whether fines of this size deter has hardly been measured.
• What we changed: The page now gives the sourced facts and marks the open question as open: the record €1.2 billion fine on Meta came to about 1% of the group's revenue the year before, roughly four days of it, against a legal ceiling of 4% (the Irish regulator's decision, which names the revenue it used); most of the largest fines are still under appeal and unpaid (the regulator reports over €4.4 billion imposed and about €20 million collected); and whether fines of this size change behaviour is contested, with little evidence either way (Lintvedt 2022). The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the administrator's review note of 15 October 2026, the research it prompted and its independent verification (as above).

• Date: 2026-10-15
• Page: Symptoms → Xenophobia & Violence; For You → Educators
• What was wrong: Xenophobia's Deeper box dated Müller and Schwarz's study 2019, the year of the working paper, while its source entry cites the published 2021 version in the Journal of the European Economic Association, which carries the outage analysis the sentence describes. Educators cited the 2022 Science Advances paper as "van der Linden et al."; its first author is Jon Roozenbeek.
• What we changed: "Müller and Schwarz (2021)"; "Roozenbeek, van der Linden et al., 2022". The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking audit of 15 October 2026 (docs/review/source-linking-audit-2026-10-15.md), checked against the papers' records.

• Date: 2026-10-15
• Page: Symptoms → Body Image; Yes, but…
• What was wrong: Body Image quotes two figures from Dove's research (one in two girls; seven in ten felt better after unfollowing), but its entry linked to the programme's homepage, where neither figure appears. Yes, but… says Aza Raskin has publicly stated he regrets inventing infinite scroll, but its entry linked to the Center for Humane Technology's homepage rather than to anything he said.
• What we changed: The Dove entry now links to the 27 April 2022 release that carries both figures and its method (an online survey of 1,551 US girls aged 10–17), still labelled brand-funded. The Raskin entry now links to the BBC News report of 3 July 2018 in which he says he "now felt guilty about it". The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: Internal review — the source-linking audit of 15 October 2026.

• Date: 2026-10-15
• Page: Laws & Regulation → Age Restrictions
• What was wrong: The timeline said that in 2023 "France passes parental consent requirements for users under 15". France did pass that law (loi n° 2023-566), but it never entered into force: after France notified it, the European Commission found that it encroached on the Digital Services Act. The page also did not say that a stricter French ban for under-15s was struck down in 2026.
• What we changed: The sentence now says the law never takes effect, and why; a new sentence says that in August 2026 the Constitutional Council struck down the under-15 ban as a limit on freedom of expression that was not appropriate, necessary and proportionate to its aim (décision n° 2026-911 DC). Both are sourced on the page. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change.
• Source: New research — the France research dossier (docs/research/FR.md, proposal 30), verified by two independent passes on 15 October 2026 (docs/review/research-verify-FR-2026-10-15.md).

• Date: 2026-10-10
• Page: Share & Support → Donate
• What was wrong: Two sentences. The page said a periodic gift "committed for five years" is deductible in full; the Belastingdienst's condition is a gift paid for at least five consecutive years, so five is the minimum, not the term. And the running-costs paragraph said "the quarterly figures below will report what was actually paid" — there are no quarterly figures on the page yet; they start once the foundation is registered.
• What we changed: "committed for at least five years", and "the quarterly figures we will publish here, once the foundation is registered, will report what was actually paid". The Dutch, Romanian and Brazilian Portuguese drafts carry the same wording.
• Source: Internal review — the Fable batch read of the Dutch edition, 29 September 2026 (issue #113).

• Date: 2026-10-10
• Page: Educate → Get Help (the resource registry, all languages)
• What was wrong: The entry for Offlimits' help line gave two numbers, "0800-0033 (free) or 020-2615275", followed by opening hours. Those hours belong to the free 0800-0033 line, the only number on the organisation's contact page. 020-2615275 appears only in the footer of its site, beside an older schedule, so the entry paired it with hours that were not its own.
• What we changed: The entry now gives 0800-0033 alone, with the hours from the contact page. Chat and mail are listed as before. The Dutch, Romanian and Brazilian Portuguese drafts carry the same change. We will re-check in January 2027.
• Source: Internal review — the independent verification pass of 10 October 2026, which re-opened hulplijn.offlimits.nl (docs/review/research-verify-NL.md).

• Date: 2026-10-10
• Page: The Attention Machine → Data Combination
• What was wrong: The page said that "in one well-documented case from 2012" an American retailer built a model predicting which customers were pregnant. 2012 is the year the New York Times Magazine published the article the page cites; the modelling it describes is older. Read quickly, the sentence dated the modelling to 2012.
• What we changed: It now says "a well-documented case, reported in 2012". The Dutch, Romanian and Brazilian Portuguese drafts carry the same wording.
• Source: Internal review — the Fable read of the Dutch and Romanian sync of the page, 10 October 2026.

• Date: 2026-09-23
• Page: Educate → Tools & Sites; Educate → Books; For Young Adults; Share & Support → Donate
• What was wrong: Six pages carried no Sources block, and writing one for each meant opening the source behind every claim on them. Six claims did not survive that. Tools described Firefox's maker as a "non-commercial, non-profit organisation": Mozilla's own governance page says Firefox is made by the Mozilla Corporation, "a wholly owned taxable subsidiary" of the non-profit Mozilla Foundation. The same page said Ecosia "anonymises searches within seven days" and that "your query and IP address do pass to Microsoft or Google": Ecosia's privacy policy says it deletes the IP address within seven days, sends it to Microsoft only in obfuscated form, and never sends it alongside the query. It also said Safari "has its own extension restrictions" and therefore offers only uBlock Origin Lite — uBlock Origin's own wiki says only that Safari support ended years ago, which is not the same thing. Books called Max Fisher a "New York Times international correspondent" (his publisher's biography says investigative reporter) and called Sander van der Linden the researcher "who developed" inoculation theory, which is William McGuire's, from 1961. Donate gave the KvK registration fee as €82, when KVK's own page has it at €85.15, and said ANBI status takes "an 8–12 week application", when the Belastingdienst says it decides within 8 weeks. For Young Adults said surveys find young adults "among the most aware of any age group" that social media is manipulative — a comparison no survey we could open supports.
• What we changed: Each of the six pages now has a Sources block and links from the claims to it. Firefox's entry names the Mozilla Corporation and its relationship to the Foundation; Ecosia's says what its policy says, and drops two claims we could not verify at all (its B Corp status and monthly public financial reports); Safari is described as not having carried the full extension for years. Fisher is "a New York Times reporter"; van der Linden is "the Cambridge researcher behind the modern prebunking studies". The KvK fee is €85 and the ANBI answer is eight weeks. The Young Adults page now states what the study it rests on actually found: in a 2025 survey of 348 US 18-to-25-year-olds, the people who knew most about algorithms were the most concerned and the least likely to act — awareness without resistance, which is the page's point.
• Source: Internal review — the sourcing pass on the six pages that had no Sources block, 23 September 2026.

• Date: 2026-09-23
• Page: Symptoms → Xenophobia & Violence; Symptoms → Elections & Democratic Harm; Symptoms → The Manosphere (and the Symptoms landing card); The Attention Machine → Your Brain; For Teens; Educate → Watch & Listen
• What was wrong: Four pages said, in the site's own voice and without a source, that the recommendation engine walks users step by step into extremism — "walks users incrementally toward more extreme content" (Xenophobia), "users are walked incrementally toward more provocative content" (Elections, offered as the mechanism behind the 64% figure), "gradual escalation keeps users engaged… each piece of slightly more extreme content performs slightly better" (Your Brain), and "you can be walked somewhere you would never have chosen to go" (For Teens). The glossary entries for "radicalisation pipeline" on For Teens and Watch & Listen said the same. The Radicalisation Pipelines page, which sets out the evidence, says the opposite: the strong claim "is not well supported", and the research is genuinely split (Ribeiro 2020 versus Hosseinmardi 2021 and Chen 2023). The Manosphere page's own metadata — its search and link-preview description, and its card on the Symptoms landing — said the worldview was "served by an algorithm that rewards radicalisation", which its body text explicitly declines to claim.
• What we changed: All six places now carry the Radicalisation page's formulation: the platform lowers the cost of finding extreme content and of finding each other, and serves extreme material preferentially to people already leaning that way; whether recommendation drives the escalation on its own is contested. Each links to the Radicalisation Pipelines page for the split evidence. The Elections page no longer glosses the 64% figure as incremental walking — that figure is a count of joins attributed to recommendation tools, not a finding about escalation. The Manosphere description and landing card now read "reached one recommended step at a time, on platforms that reward whatever holds attention".
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: About → How We Communicate; About
• What was wrong: The page that carries the site's credibility claim promised a quarterly report on comment moderation — "comments held for review, comments deleted and on what grounds" — and explained why comments are reviewed weekly rather than in real time. There is no comment system anywhere on the site and never has been. It also promised to publish "the specific headline variations tested and what each produced in terms of reach, click-through, and newsletter conversion": no headline variants are tested and there is no newsletter. And it said "First report forthcoming after launch", although the site launched in May 2026. The About page repeated the moderation promise and said the funding page is "updated quarterly".
• What we changed: One honest paragraph replaces the moderation material: the site has no comment system, correspondence arrives by email and the published engagement policy applies to it, and the same commitments apply if comments are ever opened. "Hooks used" is replaced by what the site actually counts — page views and four anonymous events, from a self-hosted cookieless analytics instance. The quarterly line now says the first report covers May–September 2026 and is published in October 2026. The section on how the site is made now says plainly that the source checks in the claim ledger were made with AI tools and re-checked by a second model, and that pages a person has read carry the date in their footer. The About page's two sentences now match.
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: FAQ; About → How We Communicate; Share & Support → Contribute; Symptoms → What We Don't Yet Know (and its Symptoms landing card); Educate → What to Do; For Teens; Yes, but…; Educate → Watch & Listen
• What was wrong: Six places promised that every claim on the site is sourced. Ten pages had no Sources block at all, and several of them named research in the body text — the Yes, but… page, which says it answers objections "with evidence, not rhetoric", cited Solove, Davison, Kahan, MI5 and the OECD in bare parentheses with nothing to follow. Watch & Listen linked "38 million households" to a #sources anchor that did not exist on the page, because its only Sources block was nested inside a glossary pop-out.
• What we changed: The promise is now the one the site keeps: every claim that carries a number, a date, a named study or a quotation is sourced on its page and tracked in the claim ledger. Real Sources blocks were added to What to Do (van der Linden and Roozenbeek, the three phone-ban studies as cited on Educators, Carter and colleagues on sleep, the OECD blog), For Teens (the Cambridge inoculation research it names twice, and Roose) and Yes, but… (Solove 2011, Davison 1983, Kahan et al. 2017, the MI5 threat update, the OECD blog, the Center for Humane Technology and Tristan Harris's 2019 Senate testimony, and the Bad News research), each linked to the primary source. Watch & Listen's Sources block was moved to the foot of the page so the anchor resolves, and gained the Netflix release behind the Adolescence claim.
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: For Lawmakers; For Parents; Educate → What to Do; Yes, but…
• What was wrong: For Lawmakers called privacy-preserving age verification "technically feasible without mass surveillance of adult users"; the Age Restrictions page calls the same question "genuinely unsolved". For Parents said of delayed smartphone access "The evidence supports it"; Age Restrictions says the outcome evidence is thin and unproven. What to Do and For Parents both said school phone bans "reliably improve focus" or "reliably cut in-school use and improve focus" — none of the three studies the site cites measured focus, and the Educators page states the fair reading correctly. What to Do and Yes, but… also described the OECD's count as "a five-fold increase in 16 months", which had drifted from the source.
• What we changed: For Lawmakers now says age verification is an active research area, not a solved one, and that the requirement and the acceptable mechanisms should be legislated separately. For Parents says the case is growing and several countries have acted — Australia has made 16 the minimum age for social media accounts — but that the outcome evidence is still thin, so this is a judgement call, not a finding. Both What to Do and For Parents now use the Educators page's wording: bans help with focus and in-school use, and probably with bullying, but are not on their own a mental-health fix. The OECD figure is now given as the counts the source gives: one country, then twelve, then twenty-five in sixteen months, of which three had a law in force by April 2026. For Lawmakers' Deeper box no longer lists "increased polarisation" as confirmed without the hedge its own body text carries ("contributed to").
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: Laws & Regulation → Courts & Liability; Laws & Regulation → Age Restrictions
• What was wrong: The Courts page set out four elements of the 2006 tobacco fraud finding and then said "All four elements are present in the social media litigation", listing three — concealment was missing. The same sentence described "features engineered to maximise addictive engagement", the framing the site's own rules forbid in its own voice, four lines above a Deeper box explaining that rule. On Age Restrictions, the observation that Australia's rule is "the first of its kind and still has only months of track record" appeared twice, once in the body and once in the Deeper box, with the second version also naming Brazil and Indonesia.
• What we changed: The Courts page now names the fourth element — internal knowledge kept from the public and from regulators — and says "features engineered to maximise engagement". The Age Restrictions body now says simply that no country's rule has been in force long enough to evaluate; the fuller account stays in the Deeper box.
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: Timeline
• What was wrong: The Timeline says it adds no claims of its own, and then made the most contested one: its 2019 entry read "Christchurch; the algorithmic radicalisation pipeline is documented." Four entries on the Machine track — 1995 "commercial web and early search engines arrive", 2009 "gamification goes mainstream", 2016 "engagement-ranked feeds become the default" and 2018 "AI recommendation systems mature" — pointed at pages that carry neither the date nor a source for it, and two of them conflicted with the destination page's own dating (How It Works dates the engagement-ranking change to 2018, not 2016). Two events the site documents in detail were absent: Cambridge Analytica (2018) and the Facebook Files (2021). The Méndez et al. 2026 source was cited without saying it is a preprint.
• What we changed: The 2019 entry now reads "Christchurch: New Zealand's Royal Commission documents the attacker's online activity." The four unsourced Machine entries were removed. Entries for Cambridge Analytica (2018) and the Facebook Files (2021) were added, each linking to the page that sources it. The Méndez citation is labelled as an arXiv preprint. The page also said "no claims of its own" twice, gave the lag argument twice and the three-track key twice; each now appears once.
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: The Attention Machine → State Actors; → Search; → Advertising & Surveillance; → Data Combination; → Under the Hood; → AI & The Next Layer; → Your Brain; → the Gamification demo
• What was wrong: State Actors said the AIVD's 2025 annual report "described online platforms as primary infrastructure for both domestic radicalisation and foreign interference" — a paraphrase an earlier correction had already retired for the BfV, and one the AIVD report does not support; the evidence base still carried the retired wording too. Search said "Research consistently shows that most users trust highly positioned results" on the strength of a single eye-tracking study, and made an unsourced accuracy claim about Google's AI Overviews although the claim ledger names the study behind it. Advertising said a profile is broadcast to "hundreds" of advertising companies per auction while the Tracking & Targeting demo said "dozens". Data Combination said of the accumulated profile "and it does not expire". Under the Hood said "in the EU the law says fingerprinting needs one [a consent prompt]" with nothing to support it, had no sources for the Belgian DPA and CJEU findings it relies on, and had lost the sentence that answers "you're doing it too". AI said instances of AI-generated content were "documented across multiple member states" and cited a paper that measures awareness, not incidents. Your Brain said "every negative word in a headline raised the click rate" without the study's design or effect size, and headed a section "Unpredictable rewards keep you hooked". The Gamification demo closed on "no conspiracy, only mechanics", not the site's line.
• What we changed: State Actors now reports what the AIVD report records — young people radicalising online and being deployed by state actors — and the evidence base notes the retired paraphrase. Search names the Tow Center for Digital Journalism's March 2025 study in the prose and in Sources (eight AI search tools; more than 60% of queries about news articles answered incorrectly), and attributes the position effect to the Cornell eye-tracking study rather than to research in general. Advertising now says "dozens", matching the demo and the ICCL source. Data Combination says "there is usually no fixed end date on it". Under the Hood cites the EDPB's Guidelines 2/2023 on Article 5(3) of the ePrivacy Directive for the consent point, adds the Belgian DPA decision and CJEU C-604/22 to its Sources, and restores the paragraph saying the site does none of this and names the analytics it does run. The AI page drops the unsupported sentence. Your Brain gives the study's design and figure — about 105,000 headline variants A/B-tested by publishers, each negative word raising the click rate by about 2.3% — and the section is now "Unpredictable rewards are the hardest to stop". The demo closes on "no conspiracy, only incentives".
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: Symptoms (landing); Symptoms → Xenophobia & Violence; → Elections & Democratic Harm; → Body Image; → What We Don't Yet Know
• What was wrong: Xenophobia said "internal documents later showed the company was aware" of the coordinated hate campaigns in Myanmar; the finding is not from leaked internal documents but from the UN Independent Fact-Finding Mission and from the human-rights impact assessment Facebook itself commissioned from BSR. Elections cited the AIVD and the BfV with no source bullets for either. Body Image gave "commonly cited clinical estimates put males at roughly one in three of those affected" with no citation. What We Don't Yet Know said the harms are "in most cases acknowledged by the platforms themselves", which is more than the site shows. The Symptoms landing said "Every major harm here was flagged… The timeline starts in 2000", while the section's own first entry is Stormfront in 1995 and not every harm on it was foreseen.
• What we changed: Xenophobia names the UN Fact-Finding Mission and the BSR assessment, and BSR's report is now in the Sources. Elections carries AIVD and BfV bullets, copied from State Actors, and names both in the prose. Body Image cites Hudson and colleagues (2007), the NCS-R lifetime-prevalence survey, for the roughly one-in-three figure. What We Don't Yet Know says "in several cases". The Symptoms standfirst now reads "Most of the harms here were flagged…" and "The story starts in the 1990s".
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: Educate → Get Help (the resource registry, all three languages)
• What was wrong: The entry for the DNSC's line 1911 said the call is free. The Directorate's own contact page says the number is reachable "at a normal rate"; the only statement that it is free came from its director-general in a radio interview.
• What we changed: The entry now says "call 1911 (standard rate), day and night": a person read the Directorate's contact page the same evening — "reachable from any network in Romania, with a standard rate, available 24/7". The number and the hours are confirmed on the Directorate's reporting platform.
• Source: dnsc.ro/contact (archived copy), pnrisc.dnsc.ro; the independent verification pass of 23 September 2026 (docs/research/RO.md).

• Date: 2026-09-23
• Page: Educate → Get Help (the resource registry, all three languages)
• What was wrong: The entry for Offlimits' help line gave its hours as Monday to Thursday 07:00–23:00, Friday 09:00–17:00 and Sunday 10:00–18:00 — the schedule printed in the organisation's site-wide footer, which has not changed since January 2026. The organisation's own hours page, updated in September 2026 together with its new free number, says Monday to Saturday 07:00–23:00 and Sunday 10:00–18:00.
• What we changed: The hours now follow the hours page: Monday to Saturday 07:00–23:00, Sunday 10:00–18:00; 0800-0033 (free) or 020-2615275. We will re-check in October 2026.
• Source: hulplijn.offlimits.nl/over-de-hulplijn/contact-en-openingstijden, re-read by the independent verification pass of 23 September 2026 (docs/research/NL.md).

• Date: 2026-09-23
• Page: Educate → Tools & Sites; What is Truth? → Science; For Educators; Yes, but…; About → Privacy; Share & Support → Donate; Share & Support → Contribute; FAQ; Contact
• What was wrong: Tools blamed uBlock Origin's Safari limitation on Google's 2025 extension changes; Safari's restriction is its own. Science said the Open Science Collaboration's 2015 paper "triggered the crisis" it in fact reported. Educators called Goodyear and colleagues (2025) "the largest English survey to date", which overstates it. Yes, but… attributed to Kahan et al. (2017) a list of techniques — conspiracy framing, authority attacks, "follow the money" — that the study did not test. Privacy said "We are a foundation", which is not yet true; the foundation is being registered. Donate offered organisations a bank transfer that does not exist yet, and listed running costs without the analytics server they are mostly spent on. Contribute said the Dutch and Romanian editions are "under native review"; no review has started, and it gave the site as about 65 pages when it now has 66. FAQ pointed "recommended tools" at the What to Do page. The Contact page listed no privacy address, although the Privacy page tells readers to write to one.
• What we changed: Tools now says Chrome and Edge allow only uBlock Origin Lite after Google's 2025 extension changes, and that Safari has its own restrictions and also offers only Lite. Science says the 2015 paper "brought the replication crisis into the open". Educators says "the largest English study to survey pupils directly (30 schools, 1,227 pupils)". Yes, but… keeps only what Kahan measured: the most numerate participants were the most polarised when the same data touched their politics. Privacy says "We are being set up as a foundation". Donate says bank transfer becomes available once the foundation's account is open, and names the self-hosted analytics server as the running cost beside the domain. Contribute says the two editions are drafted and awaiting native review, and gives the page count as 66. The FAQ links to Tools & Sites. Contact lists privacy@attentionmachine.org.
• Source: Internal review — the full content review of 23 September 2026.

• Date: 2026-09-23
• Page: Laws & Regulation → Courts & Liability; For Lawmakers; Timeline
• What was wrong: The Courts page said the case brought by more than 40 US states against Meta "runs alongside" the families' cases. On 26 August 2026, days into the trial, Meta settled with 51 US attorneys general — a minimum of $12.1 billion and up to $17.1 billion over ten years, plus default time limits, an overnight block and other design changes for under-18s under a consent judgment the court entered the same day. For Lawmakers also still said the March 2026 verdicts were "against Meta and Google", which an earlier correction had already fixed on the Courts page.
• What we changed: The Courts page describes the settlement and its terms, sourced to the consent judgment and the attorneys general's own releases, and says the families' and school districts' cases are not part of it. For Lawmakers and the Timeline carry the settlement; For Lawmakers now names the defendants of each March verdict. The claim ledger's re-check schedule now follows the whole MDL docket, not only the named cases — which is how a settlement inside the same proceedings went unseen for a month.
• Source: Internal review — the full research review of 23 September 2026.

• Date: 2026-09-22
• Page: Symptoms → Elections
• What was wrong: The page said the statement came "four days before" the 2024 election, which is the statement of 1 November 2024, but the source list linked and dated the later statement of 4 November. The official link had also stopped working.
• What we changed: The text now gives the date (1 November 2024), and the source links to that statement. The Senate Intelligence Committee report on Russia, cited as "(2019)" here and on State Actors, is now cited as the five volumes it is (2019–2020).
• Source: Internal review — the Opus review of the Dutch and Romanian editions, which read every page beside the English (September 2026).

• Date: 2026-09-22
• Page: Laws & Regulation → Courts & Liability; Laws & Regulation → Section 230
• What was wrong: Both pages read as if both March 2026 verdicts were against both Meta and Google. The New Mexico case was against Meta alone, and the design argument that sidesteps Section 230 was the heart of the Los Angeles case.
• What we changed: Courts & Liability now names the defendants in each case. Section 230 attributes the design argument to the Los Angeles verdict against Meta and Google's YouTube.
• Source: Internal review — the Opus review of the Dutch and Romanian editions (September 2026), following the correction of 21 September below.

• Date: 2026-09-21
• Page: Laws & Regulation → Courts & Liability
• What was wrong: After describing the two March 2026 verdicts, the page said "Both companies have said they will appeal both verdicts." Only the Los Angeles case involved both Meta and Google (YouTube); the New Mexico case was brought against Meta alone, so Google had nothing to appeal there.
• What we changed: The sentence now reports each case separately and brings it up to date: in Los Angeles the judge refused in June 2026 to overturn the verdict and Meta and YouTube filed appeals in July; in New Mexico the court added a $567 million abatement fund in August 2026 (a $942 million judgment in total), and Meta has said it will appeal. Neither verdict is final. Sources for each step were added to the page.
• Source: Internal review — the scheduled re-check of claims with a shelf life (issue 04.04, September 2026), against the New Mexico Department of Justice and court reporting.

• Date: 2026-09-21
• Page: The Attention Machine (landing); Laws & Regulation (landing)
• What was wrong: Both figures had been corrected on the pages themselves but not in the summary cards that link to them, which no verification pass had looked at. The Attention Machine landing still described the Search page as "Google processes 8.5 billion searches per day" — the figure replaced in the September review with Google's own, more than 5 trillion a year. The Laws landing still said "Over 25 countries have implemented or are implementing social media age restrictions since 2023"; the OECD's April 2026 count is 25 member and accession countries with restrictions in force or under consideration, of which only three had a law in force.
• What we changed: Both cards now carry the corrected figures, matching the pages they link to.
• Source: Found while preparing the Dutch translation of those landings (September 2026), which would otherwise have inherited both errors. Noted in docs/25: the claim ledger covered page prose, and these strings sit in a component's props.

• Date: 2026-09-20
• Page: About → Privacy; Share & Support → Donate
• What was wrong: The privacy page told readers that traffic was measured "with two cookieless tools: Cloudflare Web Analytics, which Cloudflare, our host, adds to every page, and Plausible, an open-source tool we host on our own server", and added that Plausible was "failing to load, so in practice only the Cloudflare figures exist". Both halves were wrong. The Plausible script was requested from a stale per-site address (/js/pa-0xUDdzxodeaF4_EiEG3yQ.js) that had been returning 404 since the site record on the analytics server changed — every page asked for it, nothing was recorded. Cloudflare Web Analytics was never switched on at all: its tag sat in the page source inside an HTML comment with a placeholder token, so it never executed, and the host was not injecting one either. The site therefore collected no analytics whatsoever, while its privacy page described two working tools. The Donate page repeated the same claim in its running-costs line.
• What we changed: The site now loads one analytics tool, the self-hosted Plausible instance, from the shared script address with data-domain rather than a hashed per-site URL that can disappear. The disabled Cloudflare Web Analytics tag was removed, along with its entries in the Content-Security-Policy, since we are not using it. The privacy page now describes one tool, says what it records, and carries a note about this correction; the Donate page matches. A new build check (scripts/check-analytics.mjs, lint CHECK 24) fetches whatever analytics URL the layout points at and fails the build on a 404, because this is precisely the kind of fault a browser and a build are both silent about. (Added 2026-09-23: the move to the shared script with data-domain was reverted the same day. The site loads the per-site script the dashboard issues, which carries the domain and the enabled captures — outbound links, downloads, form submissions — so switching one on or off is a dashboard change rather than a code one. That is deliberate. The risk this correction named — a hashed URL that disappears when the site record is recreated — is now covered by CHECK 24 itself, which fetches whatever URL the layout points at on every lint run and fails on a 4xx.)
• Source: Internal review (September 2026), reproduced against the live site and the analytics server.

The adversarial read was extended to every page the first part of the pass had not re-read (the demos, Search, State Actors, the five law pages, the What is Truth section, the Educate and Share sections, the audience pages, the objections list, the About pages). Twenty-six changes across nineteen pages; the ledger in docs/25 now covers the whole site.

• Date: 2026-09-19
• Source: Internal second-pass review (September 2026), part 2.

Repeated errors caught on other pages

• Watch & Listen, Things Worth Paying For: Adolescence "used as a discussion starter in schools across Europe" and "Cambridge University's prebunking curriculum is free" — the same two unsupported claims corrected on the Educators page earlier today; corrected here too (UK secondary schools via Into Film+; the Inoculation Science games and videos).
• Tracking & Targeting demo: "RTB as currently practised likely violates the GDPR" — the overstatement corrected on three other pages in the first pass was still on the demo page. It now says the consent RTB relies on has been found invalid and the auction runs on regardless. "Every company … retains a copy" → "can keep a copy … nothing in the system stops it", as on Advertising.
• Yes, but…, What to Do: the OECD count was "countries implementing age restrictions"; the OECD figure counts countries adopting or considering them. Corrected on both pages.

Claims stronger than the evidence, or unsourced

• State Actors, Lawmakers, Elections: the BfV report was said to identify "online platforms as primary infrastructure for radicalisation". The report calls the internet a "catalyst" (Katalysator) for radicalisation, especially of the very young. Reworded on all three pages; the North Rhine-Westphalia report's "younger and more digital" line is now sourced to the state's press release.
• Yes, but…: "Research consistently shows users are wrong about their ability to scroll past without being affected" (no such body of research is cited) and "social media became the primary information channel for most of the world's population" (overclaim) — both reworded to what is defensible.
• For Young Adults: "Research consistently finds that young adults are more aware than any other demographic" — reworded to "surveys find young adults among the most aware of any age group".
• For Teens: manosphere content "gets progressively more extreme with each recommendation" → "can get more extreme", consistent with the split evidence on the Radicalisation page.
• Digital Services Act: "For two decades, platforms were responsible for nothing" ignored the e-Commerce Directive's notice-and-takedown duty; reworded.
• Age Restrictions: "Australia's law is the only one in force long enough to study" — Brazil and Indonesia also have laws in force (OECD, April 2026); reworded on both occurrences.
• AI & The Next Layer: the Romania "Deeper" box now carries the same caveat added to the Elections page (the declassified documents were contested; a rival party had paid for an influencer campaign).
• Search: the claim that autocomplete "has been manipulated through coordinated search behaviour" now cites Cadwalladr (Observer, 4 December 2016).

Facts that had gone out of date

• Tools & Sites: uBlock Origin "works on Chrome, Firefox, Edge, and Safari". The full extension was removed from the Chrome Web Store and disabled in Chrome in 2025 (Manifest V3); Chrome, Edge and Safari run only the reduced uBlock Origin Lite; Firefox (and Brave) still run the full extension. Corrected.
• What We Don't Yet Know: the Stanford Internet Observatory was listed as a source to watch; it was wound down in 2024. Replaced with EU DisinfoLab.
• Contribute: "approximately 8,000–12,000 words per language across the full site" — the English site is roughly 45,000 words over about 65 pages. Corrected so prospective translators are not misled.

A second, adversarial pass over the whole English text (every factual sentence classified; every claim previously accepted from memory fetched from its primary source; each argument-heavy page read as a platform spokesperson, a free-speech absolutist and a sceptical researcher would read it). The full audit trail, with the status of every checked claim, is in docs/25.

• Date: 2026-09-19
• Source: Internal second-pass review (September 2026), each item verified against the primary source now named on the page.

Numbers and dates that were wrong

• Mental Health: "a US cohort study of 6,600 teenagers found roughly double the risk" for more than three hours a day. Riehm et al. (2019) report about 60% higher odds of internalising problems (aRRR 1.60 at three to six hours; 1.78 above six); "roughly double" holds only for combined internalising and externalising problems (2.01; 2.44). Both numbers now given.
• Lawmakers: the Commissariaat voor de Media's youth research was dated "February 2026". Fieldwork was March 2024 and it was published on 2 October 2024 (2,010 respondents aged 16–24, Ipsos I&O). Corrected on the page and in docs/03.
• Elections: IRA reach on "Instagram and YouTube" was "millions more". Platform testimony (31 October 2017): about 20 million on Instagram; on YouTube 18 channels and 1,108 videos with roughly 309,000 US views. Corrected and sourced.
• Watch & Listen: "Over 100 million viewers" for The Social Dilemma is a Center for Humane Technology figure. Replaced with Netflix's own: 38 million households in the first four weeks (Q3 2020 shareholder letter).
• Age Restrictions: "Over 25 countries have now implemented or are implementing" age restrictions. The OECD's April 2026 count is 25 member and accession countries with restrictions in force or under consideration; only three (Australia, Brazil, Indonesia) had a law in force. Corrected.
• Tracking & Targeting demo: "Google Analytics sits on more than half of all websites". W3Techs (2026): 43% of all websites, 78% of those running any analytics. Corrected and sourced.
• Advertising & Surveillance: "Google acquired DoubleClick in 2007" — announced April 2007, completed March 2008.
• Educators: the Bad News game "works from age 14 upward" — its makers recommend 15 and up, with a Junior version for 8–11. A "Cambridge prebunking curriculum … structured lesson plans" does not exist as described; replaced with the Inoculation Science toolkit (games, videos, a practical guide). Adolescence "widely used in secondary schools across Europe" replaced with the verifiable fact: Netflix made it free to all UK secondary schools through Into Film+ in March 2025.
• Courts: the MDL 3047 count (2,527 pending) is now sourced to the JPML report of 1 May 2026; the March 2026 verdicts are marked as under appeal (also on Lawmakers).
• State Actors: the AIVD Director-General's quotation now follows the Dutch original ("the AIVD and its predecessors … from so many sides at once, for so long").

Claims that were stronger than their evidence

• Educators, What to Do, Parents: "Research consistently supports school phone bans." The evidence is mixed: Abrahamsson (2024, Norway) finds less bullying, better grades and fewer mental-health visits for girls; Beland & Murphy (2016, England) find higher test scores; Goodyear et al. (2025, Lancet Regional Health – Europe, 30 schools) find no difference in wellbeing, sleep or overall use. Rewritten on all three pages with all three cited.
• Mental Health, Parents, Lawmakers: "effects consistently worse for girls" is now "several large studies find the association stronger for girls", with Kelly et al. (2018, UK Millennium Cohort) added as the source.
• Mental Health: Orben & Przybylski's "potatoes and glasses" comparison was about adolescents' digital technology use in general, not social media specifically; scoped. Candice Odgers' Nature review of The Anxious Generation is now cited in the Haidt-versus-Orben section. "Internal decisions to prioritise growth" replaced by what the documents show (findings played down in public; proposed fixes not adopted at scale).
• Manosphere: the lead said readers are "walked there incrementally by an algorithm"; the body that the ecosystem "reliably produces a population"; and a university news feature (Cousineau 2025) was presented as "research from Dalhousie University". The page now defers to the split evidence set out on Radicalisation Pipelines, calls Cousineau's position an argument, describes Roose's story as an illustration, and says the ecosystem "reliably reinforces a worldview".
• Elections: the Romania paragraph now records that the declassified documents were contested — a tax-authority investigation found an influencer campaign that ended up promoting Georgescu had been paid for by a rival pro-EU party — and that the annulment was criticised. The court ruling and TikTok's admission remain the confirmed facts.
• How It Works, Lawmakers: "The company chose growth" replaced with what the Wall Street Journal reported (fixes resisted because they would reduce engagement). The claim that engagement ranking amplifies divisive content was supported only by the Haugen documents; Milli et al. (2025, PNAS Nexus), a pre-registered audit of Twitter's ranking, and Brady et al. (2017, PNAS) are now cited.
• Your Brain: "content that triggers threat responses consistently outperforms content that produces calm" now cites Robertson et al. (2023, Nature Human Behaviour) and Brady et al. (2017).
• Advertising & Surveillance: "every company that receives the broadcast retains a copy" is now "can keep a copy … nothing in the system stops it" (the ICCL evidence is about broadcast, not proven retention).
• Health Misinformation: the claim that misinformation leads people to refuse cancer treatment is now sourced to Johnson et al. (2018, JNCI): alternative medicine alone, hazard ratio 2.50.
• Parents: "sudden absolute bans tend to increase secretive use" was an unsourced empirical claim; reworded as advice.

• Date: 2026-09-19
• Page: The Attention Machine → Tracking & Targeting demo; Privacy; Donate; the Demos landing
• What was wrong: (1) The live tracker test reported the Google Ads / DoubleClick pixel as "✗ blocked" for every visitor outside the United States, with or without a blocker: Google redirects that pixel to the visitor's country domain, which the page's security policy did not allow, so the browser blocked the redirect and the demo read that as protection. (2) Phase 3's "Why you" line was a fixed sentence per vulnerability; the fallback said "Professional desktop + working-hours browsing" to anyone it was picked for, including phones at night that Phase 2 had just called "Evening mobile — leisure context". (3) Phase 2 could label a device "Average / Standard" while Phase 3 called the same device "Mid-market". (4) Any language other than Dutch, German or French was labelled "English-speaking consumer" (so Romanian and Portuguese visitors were mislabelled), and every timezone in the Americas was tagged "CCPA-region". (5) The auction panel showed the animation's duration as if it were the auction's. (6) The disclosure said the tracker requests carry "nothing about you beyond the request itself"; they carry the IP address, the page address and any cookie the browser already holds for Google or X. (7) The privacy page, corrected the day before to name Plausible, was wrong the other way: the Plausible script is currently failing to load (404), while Cloudflare's own cookieless Web Analytics, injected by the host, is running. (8) The Demos landing said all three demos run "on your own data"; only one does.
• What we changed: The ad-exchange probe now uses a DoubleClick host that answers directly; the security policy and the CSP test were updated, and a new test asserts that all three probes can load under the real policy. The "Why you" line is now computed from the signals that actually qualified the visitor, and both phases derive income from one function. Languages are named or shown by code; the Americas tag no longer says CCPA. The auction timer is labelled as a slowed-down replay. The tracker disclosure on the demo page and the privacy page now says what the requests carry. The privacy and Donate pages name both analytics tools and the current fault. The Demos landing says which demo uses your data. Also: no-JavaScript fallbacks on all three demos, reduced-motion support on the tracking demo, a goal-gradient citation on the Gamification demo, and "consent not given" wording on the Data Combination demo.
• Source: Internal review of the demos (September 2026), reproduced on the live site in a clean browser.

A page-by-page check of every claim against its primary source (September 2026) found a layer of numbers and attributions that had never been verified. Each is logged below in short form; the merge request carries the diff.

• Date: 2026-09-19
• Source: Internal review (September 2026), each item verified against the primary source named.

Wrong about the site itself

• Privacy page: said analytics were Cloudflare Web Analytics and donations ran through Donorbox and Stripe, and called that "the complete list". The site loads self-hosted Plausible on every page, donations go through Ko-fi, and the Tracking & Targeting demo loads three real trackers. Rewritten to match what the site actually does, with the demo exception disclosed.
• Donate / About / Things Worth Paying For: all three disclosed affiliate income from Bookshop.org links. No such links exist and no commission has ever been earned. The Donate page also named Simple Analytics and a Buttondown newsletter (neither in use), and gave a Dutch tax example ("49% bracket … pays €51") that ignores the 1% threshold and the basic-rate cap on the gift deduction. All corrected.

Claims that contradicted the cited source

• Democracy: "The IRA's most-shared Facebook posts were about undermining confidence in electoral processes, not promoting specific candidates." The Mueller Report and Senate Volume 2, both cited on the page, found the opposite: a candidate preference plus division and voter-suppression content aimed at Black Americans. Rewritten; Senate Volume 2 added to sources.
• Science: a correction of a correction. In May 2026 we removed a "240,000-paper APS 2025 study" as non-existent and attributed a rising-replicability trend to Youyou, Yang & Uzzi (2023). The large study exists — Bogdan (2025), Advances in Methods and Practices in Psychological Science, 240,355 papers — and the PNAS paper reports no trend. Both now cited for what each actually shows, with a note on the page.
• Advertising & Surveillance: the harvest was dated "2016" (it was 2014–15); the Mueller Report and Senate report were cited for it (neither documents it — now Facebook's 2018 disclosure, the FTC's 2019 order and the ICO's 2020 letter); the ICO link went to a general page rather than the letter containing the Brexit finding; and the page said the infrastructure "was legal under the rules in place at the time" (the ICO fined Facebook the maximum £500,000 under the 1998 Act). Google's cookie reversal and its fingerprinting policy change were merged into one event; separated and sourced to the ICO's response.
• Advertising, Tracking demo, Lawmakers: "the Belgian DPA and the CJEU have found that RTB … likely violates the GDPR" overstated both rulings, which concern the consent framework and the TC String. Reworded on all three pages.
• Yes, but…: Bago, Rand & Pennycook (2022) was cited for the claim that educated people are more susceptible to targeted manipulation; the paper finds deliberation reduces conspiracy belief. Replaced with Kahan et al. (2017), Behavioural Public Policy, whose motivated-numeracy experiment does support the claim. Zahrai et al. (2022), a literature review on self-control, was cited for "design impairs the exit decision"; citation removed.

Numbers that were wrong

• Search: "8.5 billion searches per day" (Google's own figure is over 5 trillion a year); "advertising revenue $234 billion" in 2024 (Alphabet's 10-K: $264.6 billion); "AI Overviews in 30% of results by January 2025" (Semrush, the cited source: 6.49% in January, 24.61% in July 2025). All replaced and linked.
• Health Misinformation: "measles cases rose roughly 30% worldwide in 2019" — WHO reports 869,770 cases, the most since 1996, up 556% on 2016; the 30% figure was 2017 versus 2016. Corrected and sourced.
• Get-Rich-Quick: the FCA figures were misstated as "62% follow finfluencers"; the FCA's October 2024 release says 62% of 18–29s follow social media influencers and 74% of those trust their advice. Corrected and the right release linked.
• Age Restrictions: "in 2023, nearly 40% of US children aged 8–12 used social media, and 63% of UK 8–11s; a Canadian study found 86% of under-13s held accounts" — the US figure is Common Sense Media's 2021 census (38% had ever used), the Canadian figure is MediaSmarts 2021 for ages 9–11, and the UK figure could not be verified and was removed.

Attributions and dates

• How It Works, Elections, Xenophobia: the "64% of extremist group joins" figure was attributed to the 2021 Haugen disclosures; it comes from a 2016 internal presentation on German political groups reported by the Wall Street Journal in May 2020. The pages also read "users were not seeking these groups" into it, which it does not show. Re-attributed, linked, and scoped.
• Mental Health, Parents, Lawmakers: "Instagram makes body image issues worse for one in three teenage girls" now carries the qualifier in the underlying research: one in three of those who already felt bad about their bodies.
• Courts & Liability / Section 230: MDL 3047 was created in October 2022, before the states sued in 2023 (the page had the order reversed); the key Section 230 ruling was 14 November 2023, not "January 2025"; the 1990s tobacco outcome was a settlement, with the fraud liability finding in 2006. Sources added.
• Section 230: "The EFF supported the bill" — the EFF was a plaintiff against the Communications Decency Act in Reno v. ACLU; reworded.
• Xenophobia: the 2004 OSCE meeting was in Paris, not Vienna; an unidentifiable "2021 study" on social media and ethnic violence was removed; Facebook's 2018 Myanmar acknowledgment and the SPLC 2007 count are now linked.
• Consumerism: in-app shopping came to Instagram and Facebook (2019–20) before TikTok Shop, not after; the fast-fashion figures ("10% of emissions", "second-largest water consumer") are flagged as contested rather than stated as fact; the compulsive-buying claim now cites Sharif & Khanekharab (2017).
• Journalism: Johann Hari's case was plagiarism and pseudonymous attacks, not fabrication; "Woodward and Bernstein established that Nixon had authorised criminal activity" now credits the Senate hearings and the tapes for establishing it.
• Data Combination: the disputed Target "father found out from the coupons" anecdote is now flagged as disputed, with the documented modelling kept.
• Timeline / Laws landing: "the first binding response, GDPR, 23 years after hate sites were documented" paired unrelated things; reworded to "the first EU-wide rules with penalties that bite". The 2025 law entry now records the first DSA fine.
• Things Worth Paying For / Organisations: The Correspondent (closed January 2021) and First Draft (closed June 2022) removed; Information Futures Lab added.
• Citation updates: Germano, Gómez & Sobbrio is now peer-reviewed (Journal of Public Economics 255, 2026); Farooq et al. 2025, Iyengar 2019, Berridge & Robinson 1998, Riehm 2019, Izuma 2008, Valkenburg 2021, Pan 2007, the Meliorator advisory, the X fine, the White House memorandum, the NATO StratCom DSA study and the Australian Act are now linked.
• docs/03 and docs/02: still carried the "Westerwick", "240,000-paper" and Mueller-for-Cambridge-Analytica errors that the pages had already corrected, and docs/02 still linked the harvest to Brexit. Synced.

• Date: 2026-06-04
• Page: Symptoms → Xenophobia & Violence (and the Symptoms section landing)
• What was wrong: The same fact carried two different dates: the section landing said Stormfront was set up in 1994, while the page body said 1996.
• What we changed: Standardised both on 1995 — the date the Southern Poverty Law Center (the authoritative hate-group monitor, already cited on the page) gives for when Don Black set the site up (March 1995; a dial-up bulletin board preceded it in the early 1990s). docs/03 reconciled to match.
• Source: Internal review (June 2026), verified against SPLC.

• Date: 2026-06-04
• Page: Symptoms → The Manosphere
• What was wrong: A "Deeper" box grouped Cousineau (2025) among "quantitative audit studies" that test what recommender systems push to fresh accounts. Cousineau's Dalhousie work is qualitative study of the communities, not an experimental recommender audit.
• What we changed: Re-cited Cousineau as the qualitative source he is (alongside Roose 2019) and attributed the quantitative experimental-account finding to the Institute for Strategic Dialogue's YouTube audit (2022). Also linked the previously plain-text Cousineau and De Waag/Fivoor/Transfore citations to their primary sources (Dalhousie News; NOS).
• Source: Internal review (June 2026); verified Cousineau via Dalhousie News and the ISD report directly.

• Date: 2026-05-28
• Page: The Attention Machine → Search
• What was wrong: The study showing that search-result ranking of pro- vs anti-vaccination sites influences users' beliefs was attributed to "Westerwick, A. et al. (2014)." The correct authors are Allam, A., Schulz, P.J. & Nakamoto, K. (2014), published in the Journal of Medical Internet Research ("Two Experiments Manipulating Google Output"). The Westerwick search-selectivity work is a separate paper, in a different journal, on a different topic.
• What we changed: Corrected the author attribution and linked the source directly to JMIR.
• Source: Internal review (May 2026), verified against JMIR (DOI 10.2196/jmir.2642).

• Date: 2026-05-28
• Page: What is Truth? → Science
• What was wrong: The page cited "a 2025 study of 240,000 psychology papers (2004–2024)" by the "Association for Psychological Science." No such study exists. The claim is a garbled reference to Youyou, Yang & Uzzi (2023), published in PNAS, which used a machine-learning model to estimate the replicability of more than 14,000 psychology papers published 2000–2019.
• What we changed: Corrected the body text and the source to the actual study (author, year, venue, sample size) and linked it directly to PNAS.
• Source: Internal review (May 2026), verified against PNAS (DOI 10.1073/pnas.2208863120).

• Date: 2026-05-28
• Page: The Attention Machine → Advertising & Surveillance
• What was wrong: The page stated that the 87 million harvested Facebook profiles were used by Cambridge Analytica for political micro-targeting "in the Brexit referendum and the US presidential election." The UK Information Commissioner's Office, in its final 2020 report, concluded that Cambridge Analytica was not involved in the referendum beyond initial inquiries, and that the harvested Facebook data — which related to US-registered voters — could not have been used in it. Linking the harvest to Brexit was an overclaim.
• What we changed: Removed the Brexit attribution from the harvesting claim, and added a paragraph stating the ICO's finding explicitly, with a link to the ICO. The US presidential election use remains, as it is documented.
• Source: Internal review (May 2026), verified against the ICO's October 2020 report to Parliament.

• Date: 2026-05-28
• Page: Elections & Democratic Harm; Democracy; Digital Services Act (sources blocks)
• What was wrong: The Commissariaat voor de Media's May 2026 report was cited as "Naar een democratisch gezondere feed." The correct title is "Naar democratisch gezonde feeds" (subtitle: "Hoe toezicht en beleid aanbevelingsalgoritmes van sociale media kunnen verbeteren").
• What we changed: Corrected the title in all source blocks and project documents.
• Source: Internal review (May 2026), verified against cvdm.nl.