← What is Truth?

Science

The best method we have for understanding the world. How it corrects itself. How money distorts it. Why both stories matter.

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.

How science corrects itself

The replication crisis — the discovery that many published psychology and medical findings could not be reproduced — was discovered and reported by scientists, generated a global reform movement within the field, and left measurable signs that research practice has changed. The Open Science Collaboration's 2015 paper in Science brought the replication crisis into the open by replicating 100 published studies and finding many could not be reproduced. This is painful. It is also the correction mechanism working.

The field then measured itself. A 2023 study in PNAS used a machine-learning model to estimate the replication likelihood of more than 14,000 psychology papers published between 2000 and 2019 — a discipline-wide map of where the problem is worst. And a 2025 analysis of 240,000 psychology papers published between 2004 and 2024 found that, across every subfield, the typical study now reports markedly stronger statistical evidence than it did before the crisis — an indirect but consistent sign that the reforms changed practice. The field identified the problem, published about it, and changed practice. That is not a failure of science. That is science doing what it is designed to do.

An honest caveat

Influential studies continue to be cited after they have failed to replicate. The correction mechanism is real but imperfect and slow. Simine Vazire and Alex Holcombe's 2022 paper in Review of General Psychology provides an honest assessment of where self-correction works and where it does not. Science is the best method available; it is not infallible.

Manufactured doubt as strategy

Wholesale distrust of science is not a neutral position. It is a position that benefits specific actors. Naomi Oreskes and Erik Conway's Merchants of Doubt (2010) documents how the tobacco industry pioneered funding counter-research to create apparent scientific controversy where little existed — specifically to delay regulation. The fossil fuel industry used the same strategy for climate science. Anti-vaccine campaigners use some of the same tactics — skewing the science, shifting hypotheses and attacking critics. Internal tobacco documents released via litigation show this strategy was explicit and deliberate.

When someone tells you that all scientific consensus is manufactured, it is worth asking: who benefits from you believing that? Who fills the vacuum when you stop consulting verifiable sources?

When science is bought

Manufactured doubt has a quieter sibling: paying for the answer. These cases are documented from the companies' own internal papers — every one is verifiable at the links below.

Tobacco. A 1969 Brown & Williamson planning document (the "Smoking and Health Proposal") stated the strategy in one line: "Doubt is our product." Alongside the doubt campaign, the industry funded research designed to reach useful conclusions. Millions of internal documents, released through litigation, are publicly searchable in the Truth Tobacco Industry Documents archive at the University of California — you can read the strategy in the industry's own words.

Sugar. In the 1960s the Sugar Research Foundation paid Harvard nutrition scientists — roughly $50,000 in today's money, undisclosed — for a review in the New England Journal of Medicine that downplayed sugar's role in heart disease and pointed at fat instead, shaping dietary advice for decades. Documented from the foundation's internal correspondence by Kearns, Schmidt and Glantz (2016) in JAMA Internal Medicine.

Oil. Supran and Oreskes (2017) analysed four decades of ExxonMobil's communications and found a systematic gap: the company's own peer-reviewed research acknowledged human-caused climate change, while its paid advertorials promoted doubt about it.

And the pattern in aggregate. This is not a handful of anecdotes. A Cochrane review of 75 papers found that industry-sponsored drug and device research reaches conclusions favourable to the sponsor significantly more often than independently funded research on the same questions — the "funding effect", measured across the literature.

Note who exposed each case: researchers and journalists applying documentary and scientific method to the industry's own records. The lesson is not to distrust science wholesale — that is the manufactured-doubt playbook working on you. The lesson is to ask of any single convenient study: who paid for this? — and to support the structural defences that exist because of these cases: mandatory conflict-of-interest disclosure, trial pre-registration (publishing a study's plan before it starts, so an unwelcome result cannot quietly disappear), and independent public funding.

Wikipedia — an imperfect but instructive case

Wikipedia is not a primary source — the original study, document or record — but a summary of what others have published, and it says so itself: a starting point for research, not something to cite. It has been subject to editorial wars, manipulation attempts, and factual errors that persisted for years. None of this is disputed.

It is also one of the most instructive examples of the correction mechanism this section is about. Every Wikipedia article has a "Talk" tab at the top. It leads to the article's Talk page, where editors argue out in writing how the article should change — or, if nobody has yet, lets anyone start one. There is a version history where every change is recorded, and a system of flags — the best known reads "[citation needed]" — marking claims that are disputed or have no source. When Wikipedia gets something wrong, the correction process is visible. Compare this to a Google AI Overview, an algorithmically curated news feed, or a viral social media post: none of these show their corrections in the way a Talk page does, none tell you reliably when they are contested, and none keep a public record of what each person was shown.

Wikipedia's value lies in its correction transparency rather than its infallibility. In our reading, for scientific and technical subjects with active expert editor communities it is generally reliable; for politically contested subjects and current events it requires more caution. Its reference lists — the citations at the bottom of well-maintained articles — represent years of editorial work to find the original studies and documents. Start with the references, not the text.

How we know — the replication evidence, and the limits of self-correction +

The claim that science self-corrects is itself evidenced, not assumed. The 2015 Open Science Collaboration directly re-ran 100 studies. The 2023 PNAS study (Youyou, Yang & Uzzi) took a different approach: it trained a machine-learning model on features of papers known to replicate or not, then scored over 14,000 psychology papers from 2000–2019, showing that replicability varies sharply by subfield and method rather than being one field-wide number. Bogdan (2025), in an APS journal, tracked the reported p-values in 240,355 articles from 2004 to 2024 and found them growing markedly stronger over the decade after the crisis, in every subfield — not a direct measure of replicability, but the most likely signature of larger samples and pre-registration. Three different methods, one picture: the field noticed, published, and changed how it works.

A correction of our own. An earlier version of this page cited the 240,000-paper study, then removed it as non-existent and attributed a rising-replicability trend to the PNAS paper instead. Both moves were wrong: the large study exists (Bogdan 2025), and the PNAS paper does not report a trend. The corrections log records it.

The honest limit. Vazire and Holcombe (2022) document where self-correction stalls: discredited findings keep being cited for years, incentives still reward novelty over replication, and correction is slow. So "science self-corrects" is a claim about the mechanism's existence and direction, not a claim that it is fast or complete. That is precisely the contrast with systems — feeds, AI overviews, propaganda — that have no such mechanism at all.

Democracy →
Symptoms: Health Misinformation →