A recent experimental study examined how tools like ChatGPT influence thinking during writing A recent experimental study byKosmyna, N. et al. (2025), conducted at theMIT Media Lab, explores how tools like ChatGPT influence human thinking during writing tasks. The study, titled“Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,”examined how different levels of technological assistance shape cognitive engagement.
The researchers recruited participants and divided them into three groups: one using an AI assistant (LLM group), one using a search engine, and one relying only on their own thinking (Brain-only group). Across three sessions, each group used the same method. In a fourth session, roles were reversed for some participants: those who had used AI were asked to work without it (LLM-to-Brain), while those who had worked independently were introduced to AI (Brain-to-LLM). The study combined EEG brain recordings, language analysis, essay scoring, and participant interviews to understand not just performance, but underlying cognitive processes.
The results showed clear differences in how participants engaged cognitively. Brain activity, measured through neural connectivity, was strongest in the Brain-only group, moderate in the Search Engine group, and weakest in the LLM group. This suggests that as external support increased, internal cognitive engagement decreased. In parallel, language analysis revealed that essays produced with AI were more similar to each other, showing less variation in vocabulary and structure, while independently written essays were more diverse and distinct.
Participants’ experiences also reflected these differences. Those in the LLM group reported a lower sense of ownership over their essays and had more difficulty recalling or quoting what they had just written. In contrast, the Brain-only group showed strong memory recall and a clear sense that the work belonged to them. Even when AI-assisted essays scored well, they often required minimal editing and remained close to default AI-generated responses, indicating lower levels of active processing.
The fourth session provided some of the most important insights. Participants who moved from Brain-only to AI use (Brain-to-LLM) showed increased brain connectivity across multiple frequency bands, suggesting active integration of AI support with prior knowledge. They also performed well in terms of memory and structure. However, those who moved from AI use to independent writing (LLM-to-Brain) showed reduced neural engagement and did not return to the same level of cognitive activity as the original Brain-only group. Their writing also showed traces of AI-influenced vocabulary and structure, indicating a lingering effect of prior AI use.
From a clinical perspective, these findings are highly relevant. Clinical reasoning depends on active engagement, organizing information, making connections, and reflecting on decisions. Writing is one of the main ways clinicians develop and refine this reasoning. If AI reduces the need for this effort, especially early in training, it may lead to what can be described ascognitive debt: a gradual weakening of the thinking processes that support clinical judgment.
At the same time, the study suggests that AI can be beneficial when used after independent thinking has been established. The Brain-to-LLM group demonstrated that prior effortful engagement may allow clinicians or students to use AI in a more integrated and reflective way. This aligns with educational and clinical models where support tools are most effective when they build on an existing foundation rather than replace it.
These findings also echo everyday clinical practice. Therapists often emphasize the importance of active participation and reflection in patients. Similarly, clinicians themselves rely on repeated, effortful thinking to build expertise. If AI tools begin to replace rather than support this process, there may be subtle but meaningful changes in how clinicians think, remember, and make decisions.
The ethical implications are important. Clinicians remain responsible for their reasoning and documentation, even when AI is involved. The reported decrease in perceived ownership raises concerns about reduced critical engagement. If a clinician feels less connected to what they have written, they may be less likely to question it. There are also broader concerns about bias and accuracy, as AI-generated content may not always align with individual patient contexts or cultural considerations.
For researchers and students, similar risks apply. High-quality writing is not only about clarity but about understanding. If AI assists in producing text without deep engagement, there is a risk of creating work that appears strong but lacks true comprehension. Maintaining intellectual integrity requires active involvement in the thinking process, not just the final output.
Overall, this study offers an important early perspective on how AI tools like ChatGPT may shape cognition over time. For clinicians and therapists, it highlights the need for a balanced approach, one that uses AI as a support while preserving the effortful thinking that underpins clinical expertise. The goal is not to avoid these tools, but to use them in ways that strengthen, rather than replace, the cognitive processes at the heart of learning and practice.