Retraction counts measure how well fraud gets caught, not how much exists, so the hidden remainder is anyone's guess.

This explores whether retraction counts tell us how much scientific fraud exists, or only how much gets caught. The short answer is that the corpus can't give you a number, and neither can anyone else yet. It can show why retractions undercount. A retraction measures how well detection works, not how common fraud is. Several notes here point to the same problem: the tools for catching fraud are weaker than the tools for producing it.

Start with how fraud is organized. It isn't mostly lone researchers cutting corners. Paper mills, brokers and cooperating editors work as networks, with shared image banks and editor groups passing submissions to each other. When a journal loses its indexing, they move to another one Does scientific fraud operate through organized networks or individual actors?. The 2,213 articles with duplicate images were found because images can be matched against each other. Fraud that doesn't reuse images has no such trace. Moving to a new journal also resets the record, so the retraction count shows which operations got sloppy, not how large the industry is.

Generative AI makes the undetected share larger. One demonstration produced 288 complete finance papers from 96 statistically significant signals. Each came with an invented theory and fabricated citations, which turns 'hypothesizing after results are known' into an assembly line Can AI generate hundreds of fake academic papers automatically?. Human judgment won't catch this: a review of 30 studies found people spot AI-generated content at roughly chance levels Can people reliably spot content made by AI?. Readers given no provenance signals couldn't tell fluent fabrication from truth at all Can readers tell truth from fabrication without evidence signals?. Rewriting that hides authorship may also defeat AI-text detectors, although that claim hasn't been tested directly yet Do rewrites that hide authorship also fool AI detectors?.

Where detection does work, it relies on artifacts that can be checked. GPTZero flagged hundreds of possibly hallucinated citations in accepted NeurIPS 2025 papers, but how many of those flags were confirmed was never disclosed How many accepted conference papers contain hallucinated citations?. ICLR 2026 reached the same conclusion in practice. Detector scores were treated as soft signals passed to human reviewers, while confirmed fabricated references became grounds for desk rejection, because a reference either exists or it doesn't How can conferences detect and handle LLM misuse in peer review?. Enforcement goes where verification is cheap. Fraud that leaves no checkable artifact, such as selectively reported results or invented reasoning around real data, goes largely unpoliced.

The most useful idea comes from a different field. In AI training, practitioners can't tell when reward hacking starts without ground-truth labels, because the failure looks like success Can practitioners detect reward hacking without ground-truth labels?. Estimates of how common the problem is depend on how the sample was drawn, which some benchmarks don't disclose How representative is the BenchShield Trajectories labeled sample?. Scientific fraud has the same structure. Counting the undetected would require an audit sample with known ground truth, and the field rarely builds one. Until it does, the honest answer is that retractions set a lower bound, and the gap above it is probably growing.