A new study finds that asking an AI chatbot whether it would share a fake news article tells you almost nothing you didn't already know from asking whether it believes the article.
Researchers audited eight versions of large language models using 290 synthetic misinformation articles, comparing model answers to two separate survey questions: would you believe this, and would you share this. They checked those AI answers against 1,256 paired human survey responses from 317 participants to see how well each question predicted actual human behavior. An initial result looked like a problem: every model's raw "would you share" score predicted average human sharing worse than its "would you believe" score did. That gap vanished once the researchers corrected for scoring offsets, so it wasn't proof the sharing question was useless on its own. Digging further, they found credibility scores alone explained 30.8% to 72.5% of the variation in a model's own sharing answers, and adding the sharing score to a prediction already using human and model credibility improved accuracy by at most 0.69%, with most of those gains statistically indistinguishable from zero.
That matters for anyone building AI tools to screen or score content for platforms, because it suggests firing two similar sounding survey questions at a model costs twice the compute for little extra signal, not twice the insight. It is also a reminder that LLM panels standing in for human focus groups need the same incremental validity checks pollsters apply to real surveys, instead of assuming more questions automatically mean more information.
Call it the chatbot version of asking "do you like it" and "would you buy it," and getting back the same answer twice.