AI Use Makes Us Overestimate Our Cognitive Performance (www.aalto.fi)

Aalto University researchers reported in Computers in Human Behavior (Oct 27) that interacting with Large Language Models (LLMs) like ChatGPT causes widespread overconfidence: across two experiments with ~500 participants working LSAT-style logical reasoning problems, everyone who used ChatGPT substantially overestimated their performance. The study split participants into AI and non-AI groups, incentivized accurate self-monitoring, and found that although AI users often achieved higher scores, they were poorer judges of their own answers. Notably, the usual Dunning–Kruger pattern disappeared — and reversed: people who rated themselves as more AI-literate were the most overconfident. Most participants relied on a single prompt per question, copying the problem into ChatGPT and accepting the first output without verification. The result matters for AI/ML practitioners and product designers because it highlights a metacognitive gap driven by cognitive offloading: shallow interactions with LLMs reduce the cues users need to calibrate confidence and learn from mistakes, creating risks of blind trust, misinformation propagation, and workforce de‑skilling. Technical and UX remedies include designing systems that require iterative prompting, demand user explanations or justification, surface uncertainty and provenance, and provide feedback loops to foster reflection and better calibration of human-AI collaboration.