July 19, 2026
Confidently Wrong, Powered by AI
AI advice made people 3x less accurate but 2x confident, researchers found
Turns out the chatbot can make you loudly wrong — and the comments are having a field day
TLDR: Researchers found that using AI advice made people three times less accurate and far less willing to admit uncertainty, while making them feel much more sure of themselves. Commenters called it a live-action confidence scam, though some argued the real question is whether people are just as vulnerable to any polished wrong answer.
The internet has seized on a brutal new finding: give people AI advice, and they become less likely to admit they don’t know, more likely to be wrong, and somehow way more confident anyway. In the study, accuracy fell from 27% to 9%, while confidence shot up from 30% to 76%. Researchers deliberately asked the kind of movie-detail questions AI often flubs, so this wasn’t AI being secretly helpful behind the scenes — it was people getting talked into bad answers with extra swagger.
And the comments? Absolute popcorn material. One camp basically said, "Yep, this is the whole internet now." A fed-up commenter claimed advice forums have "gone to shit" because people now just paste chatbot answers and pretend it’s personal wisdom. Another user slapped a perfect label on the phenomenon: "AI assisted/induced Dunning–Kruger effect" — aka being confidently clueless with a robot hype man.
But there was also some nerdy side-eye and process drama. One commenter questioned the sourcing and complained the article’s links didn’t clearly lead back to the actual study. Another pushed a more subtle debate: is this really an AI problem, or would people get similarly misled by any slick source confidently serving up wrong answers? Still, the loudest mood was darkly amused dread. The biggest fear wasn’t just that AI is wrong sometimes — it’s that people may be learning to stop saying "I don’t know" at all, and commenters think we’re already watching that happen live across the web.
Key Points
- •The study found that with AI advice, willingness to say “I don’t know” fell from 44% to 3%, accuracy dropped from 27% to 9%, and confidence rose from 30% to 76%.
- •Researchers deliberately used questions on which AI models typically fail and selected Step 3.5 Flash because it was usually wrong on those questions.
- •Some participants who would have answered correctly without AI consulted the model and then gave incorrect answers.
- •Monetary incentives improved performance only modestly, raising ignorance admission to 8% and accuracy to 16%, still below no-AI baselines.
- •The article links the findings to earlier Wharton research on “cognitive surrender” and to concerns about AI-generated answer systems used by students and the public.