Richard Armitage is a GP and Honorary Clinical Assistant Professor at the University of Nottingham’s Academic Unit of Population and Lifespan Sciences. He is on X: @drricharmitage
Large language models (LLMs) are rapidly transforming how we work and learn, not least in medical education.1,2,3 These technologies offer revolutionising capabilities in personalising learning experiences, providing learners with immediate feedback, and democratising educational resources around the world. In the education domain, LLMs can promote independent learning, enhance student engagement, and support a variety of learning styles through adaptive content delivery.4 However, there are growing concerns about the cognitive consequences of extensive LLMs use in education.
A new paper from Massachusetts Institute of Technology – Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task – examined the cognitive effects of using LLM AI assistants like ChatGPT for essay writing tasks.5 The work revealed concerning evidence of “cognitive debt” – the accumulation of long-term cognitive costs from over-reliance on LLMs.
The authors surmise that this suggests AI assistance fundamentally restructures our cognitive architecture.
The study divided 54 participants into three groups over four sessions spanning four months: the LLM group used ChatGPT for essay writing; the Search Engine group used traditional web search; and the Brain-only group wrote essays without any tools. The researchers used EEG to monitor brain activity, conducted natural language processing analysis of the essays, and interviewed participants after each session.
With regard to neural connectivity patterns, brain connectivity was found to systematically decrease with external support levels. The Brain-only group showed the strongest, most widespread neural networks, while the LLM group exhibited up to 55% reduced connectivity compared with Brain-only participants. The authors surmise that this suggests AI assistance fundamentally restructures our cognitive architecture.
LLM users showed severe deficits in memory and essay ownership, with 83% being unable to quote from the essays they had just written, and reporting a fragmented sense of authorship of their essays.
Essays from the LLM group were also statistically homogeneous within topics, showing significantly less variation than other groups, while LLM users also used 2-3 times more named entities (such as specific facts, dates, and names) yet demonstrated less original thinking.
When LLM users tried writing without AI assistance in the fourth session, they showed weaker neural connectivity than never-AI users, while 78% were unable to quote any passage from their own essays. Conversely, when Brain-only users deployed AI tools, they showed increased neural activity and used more sophisticated prompting strategies.
The researchers coined the term “cognitive debt” to describe how LLMs spare the user mental effort in the short term but generate long-term costs including diminished critical thinking…
The researchers coined the term “cognitive debt” to describe how LLMs spare the user mental effort in the short term but generate long-term costs including diminished critical thinking, reduced creativity and independent thought, increased vulnerability to bias and manipulation, and shallow information processing.
The study suggests that while LLMs offer immediate performance benefits, early reliance on these tools might impair the development of deep cognitive skills that are essential for learning, memory formation, and intellectual independence. The researchers recommend delaying LLM integration until learners have engaged in sufficient self-driven cognitive effort to avoid the accumulation of cognitive debt.
While the results are not necessarily surprising, this paper offers the first neurophysiological evidence that LLM assistance fundamentally alters how our brains process and retain information. It raises important questions about the long-term cognitive consequences of widespread AI adoption in educational settings. While the study did not focus on medical education specifically, it is possible that the problem of cognitive debt is transferable to the over-reliance on LLMs in both undergraduate and postgraduate medical learning contexts. It might be wise, therefore, to exercise the same cautionary recommendations offered by the researchers in medical education.
References
- HC Lucas, JS Upperman, JR Robinson. A systematic review of large language models and their implications in medical education. Medical Education November 2024; 58(11): 1276-1285. DOI: 10.1111/medu.15402
- CW Safranek, AE Sidamon-Eristoff, A Gilson, et al. The Role of Large Language Models in Medical Education: Applications and Implications. JMIR Medical Education August 2023; 9: e50945. DOI: 10.2196/50945
- Y Artsi, V Sorin, E Konen, et al. Large language models for generating medical examinations: systematic review. BMC Medical Education March 2024; 24(1): 354. DOI: 10.1186/s12909-024-05239-y.
- IC Peláez-Sánchez, D Velarde-Camaqui, LD Glasserman-Morales. The impact of large language models on higher education: Exploring the connectionbetween AI and Education 4.0. Frontiers in Education 2024; 9: 1392091. DOI: 10.3389/feduc.2024.1392091
- N Kosmyna, E Hauptmann, YT Tuan, et al. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv 10 June 2025; 2506.08872v1 DOI: 10.48550/arXiv.2506.08872
