Why do catchy AI stats spread like wildfire, while the corrections limp along far behind?

This explores why a striking AI claim can circulate widely before anyone checks it, and why the correction rarely catches up. The corpus has no study that directly measures how fast hype spreads compared with corrections. It does, however, show the separate mechanisms that create the gap, and together they make a clear picture: a claim gains its credibility from how it looks, while a correction only gains credibility from being checked, and checking is slow.

The clearest case is timing. MIT publicly withdrew confidence in an AI-and-science preprint and asked arXiv to take it down, but by then the paper had already shaped the debate. The correction arrived after the damage to the discussion was done Can unreviewed preprints shape scientific debate before peer review?. Something similar happened with Deloitte's government report: fabricated quotes and fake references got through the firm's own review and reached a paying client. They only came to light afterward, and Deloitte refunded A$97,000 Can AI quality control catch fabricated citations in professional reports?. In both cases nobody verified anything until the claim was already out, so the correction was always going to arrive second.

The less obvious part is why the original claim feels trustworthy at all. Several notes suggest that people judge credibility from surface cues. In 24,000 search conversations, users preferred answers with more citations almost as strongly when the citations were irrelevant as when they were relevant Do users trust citations more when there are simply more of them?. Readers of LLM explanations come away more confident than the model's own internal probability estimates justify Why do readers trust LLM explanations more than they should?. AI-written posts collect likes because they sound thorough and sure of themselves, but they draw few replies, so they gain apparent endorsement without the back-and-forth that used to test a claim Why do AI posts get likes without inviting conversation?. A confident statistic carries all of these cues. A correction is usually narrower, more hedged and less quotable, so it carries fewer of them.

There is also a capacity problem. One note describes 'epistemic hyperinflation': AI produces claims faster than people can evaluate them, so the value of any single claim drops, the way money loses value when too much is printed Can AI generate knowledge faster than humans can evaluate it?. Verification also fails at the source. A benchmark can report reward-hacking rates of 57–73% without saying how the hacks were identified, so a number that is easy to repeat has no ground truth anyone can check How were reward hacks labeled in this benchmark study?. Even AI can make this worse: when consultants pushed back on GPT-4's output, the model argued harder for its answer instead of admitting it might be wrong Does validating AI output make models more defensive?.

The hopeful note is that checking can be scaled up too. An agentic reviewer that spends extra compute going through proofs and experiments line by line caught serious flaws in papers at STOC and ICML that human reviewers had missed Can inference scaling help reviewers catch errors humans miss?. This suggests the gap comes from where the effort goes rather than being inevitable. Producing claims is cheap and fast, while correcting them has been expensive and slow, and that cost difference may be what changes.