The Biggest AI Problem: It Makes Us More Confident — Not More Accurate
A new study from Aalto University reveals a counterintuitive insight about AI use:
When people use tools like ChatGPT, everyone overestimates how well they perform — and the most AI-literate users are the most overconfident.
The Dunning–Kruger Effect normally means low performers are the most overconfident and high performers are more accurate in judging themselves, but when using AI this pattern flips—more knowledgeable users become the most overconfident while accuracy in self-assessment declines for everyone.
Key findings in plain terms:
▶️ AI improves task performance, but users misjudge their own success
▶️ Experienced AI users are more confident, not more accurate
▶️ Most people rely on a single prompt and trust the output without checking
This leads to cognitive offloading — letting AI do the thinking, not supporting it
▶️ The real issue isn’t lack of AI skill.
It’s lack of metacognition — the ability to reflect on how and why we might be wrong.
Why this matters: Blind trust in AI risks weaker critical thinking, poor decision-making, and long-term skill erosion — even among experts.
The takeaway: AI literacy alone isn’t enough.
We need AI systems — and user habits — that force reflection, explanation, and second-guessing.
AI can make us smarter.
But without reflection, it may also make us none the wiser.
Source: “AI makes you smarter but none the wiser,” Computers in Human Behavior (2025)