Does it help readers catch AI mistakes to show them sources — even messy or incomplete ones?

This explores whether sources still help readers notice when an AI answer is wrong if those sources are imperfect. The corpus has no study that tests degraded sources on human readers directly. It does point to a surprising pattern: what helps people catch errors is having something they can check against, more than how polished the source is. In a 308-person experiment, people shown only sources caught incorrect LLM answers more often than people shown sources plus an explanation (31.8% vs. 23.1%) Do sources alone help readers catch LLM errors better than sources plus explanations?. The explanation made the answer feel more trustworthy and undid the benefit of the sources. So a plain source sometimes does more for error-catching than a fluent one that comes with a confident wrapper. Even so, under a third of wrong answers were caught in the best condition. Sources help, but they don't fix the problem.

The floor without any evidence is worse than you might expect. In one study, readers given no signal about where claims came from showed no detectable ability to tell truth from fabrication. They fell for fluent hallucinations as readily as for real facts. A display marking which claims had verified support restored their judgment Can readers tell truth from fabrication without evidence signals?. That suggests even rough provenance, meaning a sense of what's backed and what isn't, matters more than how good each source is. Systems that work over noisy material reach a similar conclusion from the other direction. A RAG system built on OCR-garbled historical newspapers works because it refuses to answer when it can't find supporting evidence. It accepts low-quality sources but won't let them stand in for an answer Can RAG systems refuse to answer without reliable evidence?.

The catch is that sources work as a signal of authority as well as evidence, and signals can be faked. LLM judges are reliably swayed by fake references and nice formatting, whatever the content says Can LLM judges be fooled by fake credentials and formatting?. Professional human review didn't catch fabricated court quotes and invented academic citations in Deloitte's government report either Can AI quality control catch fabricated citations in professional reports?. A low-quality source helps only if someone actually reads it. If people just register that a citation is there, it makes things worse, because it adds credibility without adding any checking. AI fact-checking shows the same imbalance. When the checker mislabels true headlines as false, people believe the true ones less. When it sounds uncertain about false headlines, people believe the false ones more Does AI fact-checking actually help people spot misinformation?.

There's a further point you might not expect. As models get better, their errors get harder to spot, which makes checkable evidence more valuable. Weaker models damage documents in visible ways by deleting content. Frontier models change it quietly while the text still looks intact Does model capability change how documents degrade?. A rough source you can compare against may be the only way to catch an error that reads fine. So the useful question is probably whether a source gives the reader something to check against, more than whether it's high quality. A good source that people skip because they trust it does less than a messy one they actually read.