LLM can accumulate cognitive debt

I read “Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task.” ¹ It’s useful for deciding when to use LLMs in research and writing.

Cognitive debt is when repeated reliance on external systems (like LLMs) displaces the effortful processes of independent thinking - saving effort now but costing depth, recall, and agency later.

The study finds higher neural engagement when drafting unaided. So think first, write first, then ask an LLM for critique, revise, or fact-check.

What the study did

Participants completed three essay-writing sessions in one of three conditions, plus an crossover session. Essays were written under electroencephalogram (EEG), with NLP and interviews used for analysis. N=54 completed sessions 1-3; N=18 returned for session 4. Each essay takes about 20 minutes. The 3 groups were:

1. LLM: GPT-4o only
2. Search Engine: Google with “-ai” to suppress AI answers
3. Brain-only: no tools

The study ran over 4 months.

What they found

I am interested in how should we use the technology itself from reading the paper. So I will only choose relevant conclusions. The paper offers a result of comparisons.

First of all, there is evidence indicating that the brain-only group write essays using their brain “harder”.

For example, brain-only participants showed higher activation levels in EEG and greater overall cognitive load. Writing unaided requires the brain to generate ideas, organize, and compose ideas internally. By contrast, with LLM assistance - though the essay often scored higher and were easier to write - the LLM group exerted less effort in self-monitoring. They also showed weaker memory traces and fragmented authorship.

Moreover, participants in the Brain-to-LLM group (writing first unaided, then revising with LLM) showed greater meta-cognitive engagement. They likely invested more effort in coordinating with the model and making edits to LLM’s output.

Finally, the paper finds a link between neural connectivity and behavioral quoting failures. Participants more reliant on tools encoded information more shallowly. However, meta-cognitive engagement was generally higher for the Brain-to-LLM group.

How to use LLM without “cognitive debt”

In a word, when writing something new, it’s better to draft firstly, unaided and ask an LLM to check or refine your work.

For “vibe coding” or day-to-day drafting, letting an agent like Claude run might not be the best approach for writing and learning.

Using your brain earlier helps you memorize and think deeper. Using agents or LLM engages us differently. Still, quality can be significantly improved by layering in LLM checks at the end.

1. Kosmyna, Nataliya, Eugene Hauptmann, Ye Tong Yuan, et al. “Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task.” arXiv:2506.08872. Preprint, June 10, 2025. https://doi.org/10.48550/arXiv.2506.08872. ↩