MIT Media Lab researchers ran a four-week experiment with 67 participants, asking them to judge whether news headlines and images were real or fake — sometimes with an AI chatbot helping, sometimes not.
With AI assistance, participants were 21% more accurate at spotting misinformation. The catch is that the “help” appeared to come with strings: by week four, their unassisted ability to identify fake news had dropped by 15 percentage points versus their pre-study baseline.
The most advertising-adjacent detail isn’t the human capability drop. It’s the confidence spike. Even as performance declined, a quarter of participants said they felt their detection abilities had improved.
Anku Rani, co-lead author of the paper, described the pattern as misplaced trust. People get excited about “magical” LLMs and forget they’re statistical next-token predictors with real limitations — limitations that don’t just affect outputs, but the humans leaning on them.
The study lands inside a broader pile-up of similar findings: a May study found that just 10 minutes of AI use left participants less able to solve math and SAT-style reading questions. Other research has looked at doctors, data workers, and essay writers whose independent performance or critical thinking deteriorated after relying on AI assistance.
The researchers frame this as an “AI dependency paradox”: AI boosts performance while it’s present, then skills fall below baseline when it’s removed. Valdemar Danry, the study’s other co-lead author, suggested a more Socratic AI as one possible mitigation — systems that ask guiding questions instead of handing over answers.
Read more at Fast Company.