Overview
- The MIT Media Lab presented a four-week experiment Tuesday that followed 67 participants who used an AI-backed system to judge news headline-and-image pairs.
- When users interacted with the chatbots during sessions their accuracy at spotting fake news rose by about 21 percentage points compared with unassisted judgments.
- After the AI was removed, participants’ unassisted accuracy on new items fell by roughly 15 percentage points, a decline driven mainly by a weaker ability to identify fake stories.
- Researchers call this an “AI dependency paradox” and found that conversational style mattered: AIs that ask Socratic questions or probe reasoning were linked to less skill erosion than AIs that give direct answers.
- The team cautioned the result is limited by a small set of ~50 validated items, a 67-person U.S./U.K. sample, and older models (GPT-4o plus Google Search and conversation analysis with Claude 3.5 Sonnet), and they urged replication, design changes to make AIs act as coaches, and broader AI literacy to avoid lasting deskilling.