How Fabrication Gets Found

Serious fabrication is rare and unusually damaging, and it is almost never caught by the process everyone assumes is responsible for catching it. Reviewers do not have the data. Editors rarely audit. Institutions investigating their own staff have a strong interest in not finding anything, since reputation and grant income are at stake and the employer is also the investigator.

What does catch it operates after publication, and there are three mechanisms.

The first is replication, in the ordinary working sense rather than the formal one. A laboratory tries to build on a published result and cannot reproduce the key finding. Then another does the same. For a while this circulates as private knowledge in a field (everyone knows that result does not hold) and occasionally somebody writes it down.

The second is image forensics. A small number of specialists examine published figures at scale, sometimes with software assistance, looking for duplicated regions, repeated backgrounds and splices. This has triggered a large share of recent high-profile retractions. No reviewer reading a single manuscript would notice that a band matches one published four years earlier in a different paper by a different group, and that is exactly the comparison this work makes routinely.

The third is statistical. Certain patterns are hard to fake convincingly. With integer response scales and a known sample size, only particular averages can occur, and tools exist to check whether a reported mean is among them. Terminal digits that are too evenly distributed, and variances implausibly similar across supposedly independent groups, are further signatures. None proves fabrication by itself, and each is enough to justify asking for the underlying data.

Several major cases in the past decade began with anonymous posts on public comment sites, where specialists can flag problems without risking their position, and accumulated public evidence then forced investigations that had previously gone nowhere. It is an improvised mechanism doing work the formal system was supposed to do, sustained by a handful of people willing to spend their time on it.

All of which concerns a rare phenomenon. The larger problem is not fraud, and this is the part worth taking away. Questionable research practices are common, feel defensible in isolation, and manufacture findings without anybody lying. Dropping an experimental condition that did not work. Continuing data collection until the result reaches significance and then stopping. Measuring several outcomes and reporting the one that came out. Deciding which participants to exclude after seeing the effect this has.

Each step can be justified to a colleague. Together they reliably produce publishable results from nothing at all, and the researcher's own account of events is honest throughout. The replication crisis is largely about these practices rather than about invented data, which is why the remedies are structural: preregistration fixing the analysis in advance, registered reports accepting the design before results exist, data sharing as a condition of publication.

The incentives that produced the problem are unchanged. Publication counts still decide careers, novelty is still rewarded over replication, and institutions still investigate themselves.