The Attention Machine

← For You

For Lawmakers

The evidence base, the regulatory landscape, and the policy agenda. This page is not partisan. The documented harms apply across political affiliations.

How to read this page. Tap any underlined word to see the precise term and a short definition. Expand any "Deeper" box for the evidence and contested points. The main text works on its own — you can skip both and still get the whole argument.

The evidence — what is well-established

The harms documented on this site are not contested at the level of whether they exist — they are contested at the level of precise causal weights and policy responses. The following are confirmed by peer-reviewed research, official investigations, or primary documents:

Engagement-optimised algorithms amplify emotionally provocative content — including misinformation, hate speech, and politically extreme content — because this content generates more engagement than accurate, measured alternatives. Facebook's own internal research documented this. The 2018 algorithm update increased political polarisation in the US and Italy (Germano, Gómez & Sobbrio, 2025, empirical data from both countries).

Passive social media use correlates with worse mental health outcomes in adolescents, with effects consistently stronger for girls. Facebook's own research found Instagram worsened body image issues for one in three teenage girls. The causality debate is ongoing; the direction of evidence is not.

Real-time bidding (RTB) broadcasts personal data profiles to hundreds of companies per day — 376 times per EU user, 747 times per US user (ICCL research). The Belgian DPA and CJEU have found this likely violates GDPR.

Intelligence agency corroboration

Multiple allied intelligence agencies are now saying publicly what researchers have been documenting for years. AIVD Jaarverslag 2025 (April 2026): Director-General Smit described an unprecedented threat picture, with specific concern about young people radicalising online and state actors weaponising commercial platforms. MI5 (October 2025): one in five of 232 terrorism arrests were children under 17; online radicalisation explicitly identified as the driver. BfV Verfassungsschutzbericht 2024: 57,701 extremism crimes — all-time high, up 46% — with online platforms as primary radicalisation infrastructure.

Dutch regulatory evidence

The Commissariaat voor de Media (CvdM) concluded in May 2026 that algorithmic feeds are "demonstrably risky for democracy" (aantoonbaar riskant voor de democratie) and called for better enforcement of existing European legislation. The CvdM's February 2026 youth research (2,010 respondents, Ipsos I&O) found that social media and search engines now form the primary gateway to news for most young people, with the assumption that news interest increases with age described as no longer guaranteed.

The policy agenda

Algorithmic transparency and independent audits — real researcher data access, not summarised reports. The information asymmetry between platforms (who know everything) and researchers (who know comparatively little) is itself a governance failure. The DSA mandates this; enforcement is incomplete.

Liability for harmful design — the March 2026 US verdicts against Meta and Google established that platforms can be held liable for design choices, not only for content. This principle needs legislative codification, not just litigation-by-litigation establishment.

Privacy-preserving age verification — technically feasible without mass surveillance of adult users; requires legislative specificity about the mechanisms that will and will not be acceptable.

Mandatory algorithmic impact assessments — before deployment, not after harm is documented. The DSA moves in this direction; the approach needs global adoption.

Platform funding for independent research — analogous to requirements imposed on the tobacco industry to fund cessation research. Platforms have the data; researchers need access to it.

Deeper — reading this evidence base: confirmed, contested, and what that means for legislating +

For policy purposes the evidence here sorts into three tiers, and conflating them weakens any resulting law. Confirmed (peer-reviewed, official investigation, or primary document): that engagement-ranked systems amplify provocative content; that the 2018 Facebook change increased polarisation in two countries; the RTB data-broadcast figures; the intelligence-agency assessments; the March 2026 design-liability verdicts. Contested: the precise causal weight of social media in the youth mental-health decline. Genuinely unknown: the long-term developmental effects, and the impact of generative AI on elections.

The implication for drafting. The robust legislative targets are the confirmed, structural facts — transparency, auditability, design liability, data access — because they do not depend on winning the contested causation debate. A statute that rests its case on "social media causes the mental-health crisis" can be litigated to a standstill on the science; one that rests on "these systems are opaque, their design choices are the company's own conduct, and independent verification is currently impossible" stands on confirmed ground. Legislate the asymmetry and the design, not the contested effect size.

Laws & Regulation — the full regulatory landscape →