An Era of Cognitive Dependence?

Generative artificial intelligence, and more specifically ChatGPT, has established itself in record time as one of the most widely used tools in educational, professional, and entrepreneurial spheres. But what is it really doing to our brains? The MIT study entitled “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task” provides unprecedented insight into the neurocognitive effects associated with prolonged use of LLMs. Its finding is unequivocal: increased dependence on these tools leads to “cognitive debt,” a measurable decrease in brain activity related to memory, attention, and autonomous reasoning ability.

A Pioneering Study: Rigorous Scientific Protocol

MIT researchers followed 54 participants over four months, divided into three groups:

  • Group 1: LLM only (ChatGPT),
  • Group 2: search engine (without AI),
  • Group 3: brain alone, without digital tools.

Each participant had to write argumentative essays based on exam topics (SAT-type). Their brain activity was recorded via EEG, while their outputs were evaluated by human teachers and an expert AI. A final session reversed the roles: those who had used ChatGPT had to write without tools, and vice versa. This design made it possible to observe both immediate and cumulative effects of AI use.

Results 1 — Fewer Neural Connections with ChatGPT

The EEG analysis reveals a direct correlation between the type of digital assistance used and brain engagement:

  • The Brain-only group shows the greatest neuronal activity, extensively mobilizing the alpha, beta, delta, and theta bands.
  • The search engine group shows intermediate engagement, particularly in visual-executive networks.
  • The ChatGPT group displays significant under-activation, with reduced connectivity across all frequency bands.

In other words, the more the user relies on AI, the less their brain works. This phenomenon is all the more pronounced as use is repeated: participants exposed to ChatGPT during several sessions show a lasting drop in neural connectivity even when they stop using it.

Results 2 – Failing Memory and Weakened Sense of Ownership

Immediate memory tests show a striking result: 83% of ChatGPT users are unable to quote a sentence from their own essay, compared to 11% in other groups. Qualitative analyses confirm a weak sense of ownership: half of LLM users say they don’t feel like “authors” of their texts, despite high overall satisfaction.

In post-task interviews, several participants acknowledge cognitive “disengagement”:

“I don’t even remember what I wrote, it was ChatGPT doing everything,” one participant testifies.

Results 3 – More Homogeneous Essays, but Less Creative

Linguistic analysis of the texts reveals lexical homogenization in the LLM group: the same phrases, the same proper nouns, the same n-grams appear in the majority of texts, regardless of the participant. ChatGPT tends to provide pre-formatted narrative structures, with few thematic variations, while Brain-only group essays show much greater lexical and conceptual diversity.

Even when texts produced by ChatGPT are well-rated, their content proves conventional, impersonal, and sometimes disconnected from the author’s experience. The study thus shows that the quality perceived by the teacher or by an AI judge does not necessarily reflect the cognitive quality of the writing process.

Analysis – Cognitive Debt: A Measurable Phenomenon

MIT researchers use the concept of “cognitive debt” to designate the progressive reduction in mental effort required by a task, as one delegates thinking to an automated assistant. This debt manifests itself through:

  • A loss of cognitive autonomy,
  • Atrophy of neural circuits related to analytical thinking,
  • Weakened working memory,
  • Decreased motivation and ownership.

The study also highlights a reverse adaptation phenomenon: those who used the brain alone for three sessions, then ChatGPT in the last one, show stronger brain connectivity than those who took the reverse path. This suggests that initial learning without AI partially protects against cognitive atrophy effects.

The Illusion of Performance: When Substance Fades Behind Form

One of the major lessons of this study is the gap between perceived performance and cognitive reality. Texts produced with ChatGPT are more fluid, structured, without apparent errors. Yet this ease conceals a weakening of critical thinking, memorization, and the ability to argue originally. AI thus becomes a competence simulator, but not an intelligence catalyst.

Implications for Education – From Assistance to Substitution

The authors warn about educational consequences in a school or university context. Intensive use of ChatGPT could lead to massive unlearning of basic skills, particularly among students with low self-esteem or struggling. Even more worrying, the study highlights decreased motivation among those who rely excessively on AI: less effort, less intrinsic satisfaction, less perseverance.

Pedagogies centered on autonomy, argumentation, and creativity therefore risk clashing with instrumentalized AI use that transforms learning into passive automation.

Business Impacts – Short-Term Performance, Long-Term Vulnerabilities

In business, LLMs are praised for their immediate productivity gains. But what is the value of expertise that is not anchored in the employee’s memory? This study directly questions the promises of AI assistants in training contexts, report writing, or strategic content generation.

If competence becomes dependent on an external tool, teams’ cognitive capital becomes fragile. The company finds itself exposed to a double risk:

  • Loss of cognitive resilience in case of unavailability or evolution of tools,
  • Standardization of generated content, harming innovation, differentiation, and critical thinking.

Regulatory Issues – For AI That Stimulates Rather Than Replaces

The MIT study suggests concrete avenues: favoring a “dialogical” use of AI, where humans remain masters of reasoning, asking questions, selecting ideas, confronting answers. It also advocates for qualitative measurement criteria, focused on actual cognitive effort, not just textual output.

This calls for a paradigm shift in companies and educational institutions: integrating AI as an intelligent tutor, not as a cognitive substitute. Training in meta-cognition becomes a priority.

 

Toward Virtuous Hybridization?

Rather than opposing human and AI, MIT authors propose a third way: reasoned cogno-design, in which AI acts as an extension of human reasoning, without short-circuiting its mechanisms. This implies:

  • Transparent software design,
  • A logic of human-machine collaboration,
  • An ethics of continuous learning.

This model, however, remains to be invented. And for now, the effects documented by this study call for caution.

ChatGPT: Prodigious Tool or Invisible Threat?

The MIT study does not demonize ChatGPT. Rather, it offers us a lucid reading of its invisible effects. If we don’t question its uses, generative AI could well become the most sophisticated tool ever invented… for unlearning.

 

Main References