Your Brain on ChatGPT

As we all continue to explore the possibilities of integrating meaningful AI into teaching and learning, this week I would like to share a preprint entitled, “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” by  MIT researchers Kosmyna, et al. (June 2025). At first the title may appear to reveal an unexpected, negative effect of using AI for learning. However, with reference to how humans process information, this study validates the research that we have collected for many years. 

The study divided participants into three groups:

1. LLM group, who used ChatGPT to help write essays;

2. Search Engine group, who used traditional web search; and

3. Brain-only group, who relied solely on internal knowledge.

Across multiple sessions, electroencephalography (EEG) data revealed that brain connectivity scaled inversely with technological support. Those in the Brain-only group demonstrated the strongest neural engagement, particularly across alpha and beta networks related to attention and working memory. The LLM group, in contrast, showed under-engagement, suggesting they were outsourcing not just the writing but also the cognitive processes that support learning.

By session 4 — when the LLM group was asked to write without AI — their diminished neural activity was pronounced. They struggled to recall their own work, and scored lower on essays. Meanwhile, the Brain-to-LLM participants showed strong recall and reactivation of prior neural pathways, suggesting that foundational cognitive engagement builds transferability — even when switching to AI support later.

Cramming, Offloading, and the Illusion of Learning - This study confirms what decades of learning science have already taught us: Cognition requires effort, and effort leads to retention. For years, researchers like Bjork & Bjork (1992) and Dunlosky et al. (2013) have argued that desirable difficulties — retrieval practice, active engagement, elaboration — build stronger long term memory (LTM). Conversely, passive review, cramming, and over-reliance on external aids lead to fragile, short-term learning.

What Your Brain on ChatGPT reveals is that AI use can recreate the illusion of learning — similar to last-minute studying — where polished output masks shallow processing. The LLM group produced coherent essays, but their interviews and EEG results showed low ownership, weak memory, and minimal neural activation. This aligns with prior cognitive models: when learners aren’t challenged to retrieve, reflect, and reorganize knowledge, they don’t encode it deeply. The tool may “do the job,” but the learner doesn’t reap the cognitive rewards.

This doesn’t mean AI has no place in learning. Faculty might rethink prompt design, task framing, and authentic assessment models. Rather than allowing AI to replace cognitive labor, educators could use it to augment:

• Prompt students to analyze, critique, or revise AI outputs.

• Require reflective annotations on how they used AI and what they learned from it.

• Encourage pre-writing brainstorming without AI, followed by AI-assisted revision — engaging both effortful retrieval and scaffolded feedback.

A major takeaway from this study might be about pedagogical intentionality. LLMs are tools, not teachers. When used without guidance, they can undermine the cognitive processes that learning is designed to foster. But when paired with active pedagogy and awareness of cognitive science, they can become powerful supports.

References

Kosmyna, et al. (June 2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. Preprint. Retrieved at https://arxiv.org/pdf/2506.08872