The most widely published sentence in technology today may also be the most ignored: AI can make mistakes. It sits in small gray type beneath nearly every chatbot and is repeated millions of times a day, yet most of us stopped seeing it long ago. So what do people actually know about AI's mistakes, and what do they do when they encounter one?
New research from the Ad Council Research Institute (ACRI) and Google offers some of the first nationally representative answers, and they point to a simple conclusion. Americans are not naive about AI's flaws. What they want is not another reminder that the technology is imperfect, but clear guidance on what to do about it.
The ACRI and Google paper, Why and How People Check Generative AI Output for Mistakes, was accepted to the Ninth AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026) in Malmö, Sweden. AIES is one of the field's leading peer-reviewed venues for research on how AI shapes society, bringing together computer scientists, ethicists, legal scholars, and social scientists.
We surveyed 1,503 U.S. adults in May 2025, with a sample matched to Census projections on age, gender, race and ethnicity, region, and household income. The survey included both users and non-users of AI. Respondents described in their own words whether they checked AI results and how, and then reviewed disclaimers, highlighting the language they found helpful and the language they did not. Here is what we learned.
The public already knows AI gets things wrong.
Seven in ten Americans (70%) say generative AI makes mistakes at least sometimes. Only 2% believe it never does. A majority believe AI at least sometimes gets facts wrong, invents things that don't exist, contradicts expert consensus, and ignores the instructions it was given.
Users and non-users recognize AI's fallibility at broadly similar rates, and the people who know the technology best are the most clear-eyed about it: respondents who are very familiar with generative AI were 2.64 times as likely as unfamiliar respondents to say it makes mistakes often. Trust follows the same pattern. Only 11% of Americans find AI results very trustworthy, while the rest hold at least some measure of doubt.
Most people check, and the reasons they don't are telling.
Three in four respondents who answered the question said they check AI output at least some of the time, and 13% said they always do. The most common method is cross-referencing: a quick search, a look at the sources the tool cites, or a comparison with a trusted website. Some compare answers across multiple AI tools or ask a knowledgeable person. The motivation is often personal accountability. As one respondent put it, "I do not want to send something wrong out to a client."
The quarter who rarely or never checks provides some interesting insights as to why. Some use AI only for low-stakes tasks. Others have come to trust it over time, or believe it must be right because it "has access to the entire internet." Many simply see checking as a contradiction, since they turned to AI to save time in the first place. And among those who do check, about one in nine wait until something looks suspicious, a habit that can let the subtle, plausible-sounding errors slip through precisely because they don't look wrong.
What works in a disclaimer, and what doesn’t.
The disclaimer highlighter exercise revealed a sharp divide between warnings and guidance. Generic cautions such as "AI can make mistakes" or "it's not without its risks" earned only lukewarm reactions, and language suggesting answers "might be changing daily" was actively disliked. In contrast, direct calls to action, such as verifying results or double-checking before sharing, were among the most favored phrases in every statement tested. So was any invitation to help improve the product by leaving feedback.
People also wanted more than reassurance. When asked what else they would want to know, respondents asked how exactly they should check results, how often AI is actually wrong, where its information comes from, and what its environmental costs are. One captured the tension plainly: having to check seems to defeat the purpose, so tell me how to do it quickly and easily.
As AI moves from answering questions to acting on our behalf, the stakes of getting this right will only rise. The good news in this research is that Americans are already doing much of the work themselves. The task for everyone who builds, deploys, or communicates about AI is to meet them where they are, with honesty about limits and real help in navigating them.
The full paper is available now on arXiv and will appear in the AIES 2026 proceedings.
Photo by Kindel Media/Pexels