Metacognition: knowing what you know "AI brain rot" and cognitive offloading
With AI, people scored better on reasoning puzzles but overestimated how well they did
In short: People who used AI to solve logical reasoning problems from the LSAT scored a little better than a norm group but overestimated their own score by even more. People with higher AI literacy were more confident and less accurate about how they had done.
| Paper | AI makes you smarter but none the wiser: The disconnect between performance and metacognition |
| Authors | Daniela Fernandes, Steeven Villa, Salla Nicholls, Otso Haavisto, Daniel Buschek, Albrecht Schmidt, Thomas Kosch, Chenxinran Shen, Robin Welsch |
| Published | February 2026 · Computers in Human Behavior 175 (February 2026 issue; online from October 2025) |
| Study type | Two online experiments · Peer-reviewed |
| What we read | Abstract only. Summary based on the published abstract only. |
| Summarized |
What they did
Participants solved 20 LSAT logical reasoning problems with access to an AI and then estimated how many they had got right. The researchers compared performance with a norm population, compared estimates with actual scores, and used a computational model to look at individual differences. A second study repeated the design.
Who took part
Study 1: 246 participants. Study 2: 452 participants.
What they found
- With AI, performance was about three points higher than the norm population.
- Participants overestimated their performance by about four points.
- Higher AI literacy went with lower accuracy in judging one's own performance.
- The usual Dunning-Kruger pattern on this task (low performers overestimating most) disappeared with AI use: overestimation was more even.
- Study 2 replicated the findings.
Effect size: About +3 points in performance and about +4 points of overestimation (Study 1, out of 20 problems).
Limitations
- Adults on logic puzzles, not students learning a subject.
- Measures performance with AI, not learning afterwards.
- We read only the abstract.
What this means for learners and parents
- When AI helps, it is harder to know what you yourself can do. Test yourself without it.
- Feeling confident after using AI is not good evidence of understanding.
- Explaining an answer back in your own words, without the screen, is a quick reality check.
Put it into practice (free tools)
Citation
Daniela Fernandes, Steeven Villa, Salla Nicholls, Otso Haavisto, Daniel Buschek, Albrecht Schmidt, Thomas Kosch, Chenxinran Shen, Robin Welsch (2026). AI makes you smarter but none the wiser: The disconnect between performance and metacognition. Computers in Human Behavior 175 (February 2026 issue; online from October 2025). https://doi.org/10.1016/j.chb.2025.108779
Related summaries
- Unrestricted ChatGPT help raised practice scores but lowered exam scores · Hamsa Bastani et al., 2025 · Preregistered randomized controlled trial
- "Your Brain on ChatGPT": what the MIT essay-writing study did and didn't show · Nataliya Kosmyna et al., 2025 · Lab experiment with EEG
- ChatGPT made research feel easier but students' reasoning was weaker than with Google · Matthias Stadler et al., 2024 · Randomized experiment
- The classic study: testing yourself beats rereading after a week · Henry L. Roediger III et al., 2006 · Lab experiments