LLMs are bad for learning. When former LLM users are asked to write solo, they show clear evidence of cognitive debt and skill atrophy.
This EEG study tracked brain activity during writing tasks across three experimental groups: Tool-free, LLM-assisted, and Search engine-supported.
Writing on our own stimulates maximum neural activity as it forces memory retrieval, idea generation, linguistic formulation, and self-monitoring.
With an LLM, neural activity plummets. Because the model structures and generates the draft, our brains delegate the heavy lifting. We stop synchronizing thoughts and maintaining a cohesive plan.
The long-term tradeoff is even more striking. Over four months, the LLM group performed worst across the board: neural connectivity, writing originality, and retention. The search engine group performed somewhere in the middle.
LLMs are not tools in the same way as calculators, the Internet, or search engines. None had the autonomy to hijack the mechanics of thinking and planning as LLMs do. I’ve been making this distinction since early 2024 in my talk, “Mind the AI Gap,” and making a case for pre-prompt work while learning.
Sources: MIT Study – “Mind the AI Gap” Talk