How a Claim Earns Trust

An institute that asks what human beings and institutions should become as machines grow more capable owes its reader an account of its own method, not only its subject. This essay is, in that sense, unlike the three that precede it. It does not argue primarily about artificial intelligence. It argues about how any claim — including the ones this Institute makes — comes to deserve belief, and it takes that argument from a field that has already been forced to answer the question under pressure: experimental psychology, over the past two decades.

A field discovers it cannot trust its own literature

Ioannidis (2005) made the theoretical case before the empirical one existed: given small sample sizes, flexibility in study design and analysis, financial and other interests, and a research culture that rewards novel, striking findings over careful null results, most published research findings in many fields were more likely than not to be false — not through fraud, in the typical case, but through the ordinary operation of incentives that reward being interesting over being right.

A decade later, the Open Science Collaboration (2015) supplied the empirical test, for psychology specifically. Researchers attempted to replicate 100 studies published in three leading psychology journals, using the original materials and high-powered designs wherever possible. Ninety-seven percent of the original studies had reported statistically significant results. Thirty-six percent of the replications did. Where an effect did replicate, its magnitude was, on average, roughly half the size originally reported. The finding was not that psychology’s published literature was fraudulent. It was that a field’s collective confidence in its own findings had been running well ahead of what those findings could actually support — and that this had happened without any single actor doing anything most researchers, at the time, would have called wrong.

What actually fixed it

The response that followed did not consist of exhortations to greater rigor or improved peer review in the abstract. It consisted of specific, adopted mechanisms, each sharing one structural feature: a public commitment, made before the outcome of a study is known, to what result would count as support for a hypothesis and what would count against it.

Chambers (2013) introduced Registered Reports at the journal Cortex: a submission format in which a study’s introduction, hypotheses, and analysis plan are peer reviewed and can receive in-principle acceptance for publication before data collection begins — meaning the decision to publish is made on the quality of the question and the method, not on whether the result turned out to be interesting. The format has since been adopted, in some form, by several hundred journals. Nosek, Ebersole, DeHaven, and Mellor (2018) documented the broader adoption of pre-registration across the discipline as part of what they term the “preregistration revolution,” and Nosek and colleagues (2015), writing in Science as the Center for Open Science, proposed the Transparency and Openness Promotion guidelines now used by many journals to specify, at the level of editorial policy rather than individual author discretion, what data, materials, and pre-registration a published study is expected to disclose.

The common mechanism across all three is worth stating precisely, because it is the part that transfers. None of these reforms rests on trusting researchers to be more honest. Each removes, structurally, the opportunity to decide after the fact what the hypothesis was, which analyses to report, or what would count as confirmation — by requiring that commitment be made publicly, in advance, while the outcome is still unknown to everyone, including the person making the claim.

The same vulnerability, a faster mechanism

The vulnerability Ioannidis identified was structural before it was ever about any particular technology: credibility asserted or defended only after an outcome is known, by the party with the greatest interest in that outcome being believed. Generative artificial intelligence does not create this vulnerability. It lowers the cost of exploiting it, at a scale the reforms above were never built to anticipate — a separate concern this publication has examined directly, in the growth of AI-assisted paper mills producing plausible-seeming research output faster than existing review processes can evaluate it. The reform lineage that rebuilt credibility in experimental psychology is the relevant one here not because artificial intelligence resembles a psychology experiment, but because the underlying failure is the same failure: a claim protected from disconfirmation by being evaluated only in retrospect.

This is the method this Institute has adopted, and the reason it takes the form it does. The Standards of Inquiry set out in the Institute’s Charter and Editorial Charter — evidence, historical seriousness, institutional realism, pluralism, revision, intellectual independence — are published once, in advance, as the standard against which any claim this Institute makes, including this essay’s, may be judged; they are not assembled after a claim has already been made to justify it. The Disputation Protocol asks the same discipline of a live exchange that Registered Reports ask of a study: before the session, each side states what evidence or argument would change its position, so that the outcome cannot be redefined after the fact to declare victory regardless of what occurred. Governed Intelligence, at the institutional level this publication studies in others, asks that delegated machine execution remain observable and revisable rather than judged only by whether its output looks right after the work is done — the same principle, applied to institutions using AI rather than to researchers studying human subjects.

Four theses

This essay’s claims are subject to challenge under the terms of the Institute’s Disputation Protocol. Editorial disclosure, including this publication’s AI review process, appears in the closing section of this edition.

References

Chambers, C. D. (2013). Registered Reports: A new publishing initiative at Cortex. Cortex, 49(3), 609–610.

Ioannidis, J. P. A. (2005). Why most published research findings are false. PLoS Medicine, 2(8), e124.

Nosek, B. A., Alter, G., Banks, G. C., Borsboom, D., Bowman, S. D., Breckler, S. J., … Yarkoni, T. (2015). Promoting an open research culture. Science, 348(6242), 1422–1425.

Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11), 2600–2606.

Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349(6251), aac4716.