AI and Human Thought: A Worrying Erosion

The Impact of AI on Human Thinking

Researchers from MIT, Oxford, and Carnegie Mellon conducted a study involving 1,222 participants faced with math and text comprehension exercises. The results show that those who used AI produced more incorrect answers than those who worked without assistance. This phenomenon was observed across three distinct samples.

Daron Acemoglu, the 2024 Nobel Prize winner in Economics, emphasizes that while AI can improve the quality of decisions for those trained before its advent, it tends to reduce incentives for learning, which are essential for long-term collective knowledge. Participants with access to AI also showed a tendency to give up more quickly when faced with difficulties, illustrating a form of intellectual laziness.

Overreliance on AI

A study conducted by Microsoft Research and Carnegie Mellon on 319 knowledge workers revealed that overreliance on AI leads to less verification of the results produced by the machine. In contrast, those who trust their own skills continue to maintain a critical perspective.

The researchers highlight the irony of automation: by mechanizing routine tasks, the user loses opportunities to exercise their judgment. When an unexpected situation arises, cognitive abilities are weakened, and the user merely checks and integrates the response provided by AI.

Intensive users juggling multiple AI tools commit 39% more major errors, according to a Boston Consulting Group survey of 1,488 American professionals. This phenomenon, dubbed "AI Brain Fry," manifests as cognitive fog and headaches.

Effects on High School Students' Learning

In 2024, an experiment conducted on high school students using ChatGPT for math practice showed that they solved 48% more problems correctly than those without AI. However, during a final test without assistance, their results were 17% lower.

A third group, using a modified version of ChatGPT that provided hints without giving the solution, solved 127% more problems during training but showed no gains during the final test compared to the unassisted group.

Brain Analysis and Recommendations

In 2025, at the MIT Media Lab, Nataliya Kosmyna used electroencephalography to measure the brain engagement of participants using ChatGPT. The results showed lower brain engagement compared to those using Google or working alone. The alpha and theta waves, associated with deep memory, were particularly low, indicating that little information was being integrated into memory.

Nataliya Kosmyna recommended developing the brain in an analog manner before exposing individuals to AI, and learning to use these tools without delegating the necessary mental work for learning.