AI Writes for Us and the Brain Conserves Energy

Measuring the brain activity of someone writing with the help of AI and someone writing alone produces results that should be troubling far beyond neuroscience. The study conducted in 2025 by Nataliya Kosmyna and her colleagues at the MIT Media Lab shows that regular users of AI writing assistants display significantly lower measured brain activity than those who write without assistance. The brain, confronted with a tool that does the work for it, economizes. The question is what it loses in economizing.

This is not an alert about technology. It is an alert about usage. And the distinction is decisive.

The Essential Points

  • The Kosmyna et al. study (MIT Media Lab, 2025) documents an accumulation of measurable “cognitive debt” among regular users of AI for writing, visible in brain activity signals.
  • Neuroscience distinguishes between two types of cognitive effort: production effort (generating ideas, structuring, formulating) and validation effort (reading, correcting, approving). AI tends to substitute the former with the latter.
  • Pedagogy is where the question is decided: initial experiments conducted in American and European institutions show that the framing of usage directly influences cognitive gains or losses.
  • The issue on a generational horizon: if generative AI becomes the dominant writing mode for 15-25 year-olds, the effects on structured reasoning capacities will only be measurable with a delay of ten to fifteen years.

What the Brain Does When It No Longer Writes

Writing is not an act of transcription. It is an act of thinking. Writing a sentence forces you to choose between formulations, to clarify a vague intention, to discover through writing what you actually think. Research in cognitive psychology has documented this process since the 1980s: the constraint of the written sentence acts as a tool for conceptual clarification. We do not know what we think until we write it.

What the Kosmyna et al. study shows is the flip side of this coin. When AI produces the text and the user validates or corrects it, the type of cognitive engagement changes in nature. The brain shifts from a production mode to an evaluation mode. These two modes mobilize distinct networks. The production effort, particularly in the prefrontal zones linked to planning and formulation, is what researchers observe in decline among regular users of AI assistance. They characterize this effect as “cognitive debt”: a form of functional atrophy from disuse, analogous to what is observed in individuals who cease to practice a physical skill.

The mechanism is not mysterious. The brain is an organ of energy optimization. When an external tool takes over a task, it reduces the effort allocated to that task. This is an advantage in many contexts. It is a problem when the delegated task is precisely the one that forges the capacity.

The Pharmakon: The Same Molecule That Heals Can Poison

Anna Alombert, a philosopher specializing in technology, uses the concept of pharmakon, borrowed from Bernard Stiegler, to think about generative AI. A pharmakon is both remedy and poison depending on the dose, usage, and context. The calculator is a pharmakon: it frees the mind from repetitive calculations and allows one to go further in mathematical reasoning, but it atrophies mental calculation capacity if it replaces learning arithmetic itself in children.

Generative AI applied to writing operates on the same pattern. Used by a professional who already masters writing and who uses it to save time on standardized tasks, it frees up cognitive bandwidth for more complex tasks. Used by a student who is learning to structure an argument, it short-circuits precisely the learning process they must go through. The same tool, the same gesture, two opposite effects depending on who uses it and why.

This framework allows us to move beyond the sterile debate between those advocating for a ban (AI degrades brains) and defenders of total adoption (AI enhances capacities). Both are right depending on context. The productive question is: in what uses does AI enhance, and in what uses does it dispossess?

The Concurrent Reading: What Daron Acemoglu’s Data Adds

A different reading exists, and it deserves to be taken seriously. In his recent work on technology and power, Daron Acemoglu argues that the real effects of tools on human capacities depend less on the tools themselves than on the structures in which they are deployed. The argument applies here: the question is not “Does AI cognitively dispossess?” but “Who controls the framing of uses, and in what interest?”

Large generative AI platforms have a business model based on engagement and retention. A user who delegates more to the tool is a more dependent user, thus more loyal. The architecture of these tools is not neutral: AI systematically proposes a complete answer before the user has even formulated their thought. It does not propose cognitive scaffolding that would help the user build their own answer. It proposes a finished answer. This design produces exactly the effect that Kosmyna et al. measure: the validation mode replaces the production mode.

This does not make the tool bad. It means that the cognitive effort preserved or not depends not on the technology but on the design of the interaction. And the design can be something other than what it is today.

What Pedagogy Can Do That Prohibition Cannot

The first documented educational experiments on this subject provide concrete indications. American and European institutions have tested contrasting approaches since 2023. Those that simply banned AI obtained massive and undisclosed workarounds. Those that integrated AI with explicit protocols produced more interesting results.

Khan Academy developed Khanmigo, an AI assistant that deliberately refuses to give answers directly. It asks questions, guides, forces the student to formulate their reasoning before validating or correcting. The architecture of the interaction is designed to preserve production effort. The first evaluations, still preliminary, constitute a promising lead regarding this mode of usage’s capacity to limit the effects documented by Kosmyna et al., even though robust data remain to be built.

In France, the 2024 Bergounioux-Thiollière report on digital technology in schools recommended a similar approach: defining AI usage zones according to age and targeted competency. The idea is that cognitive delegation is acceptable when the competency is already acquired, and that it is counterproductive when the competency is being acquired. A high school student who has learned to structure a plan can use AI to accelerate writing a standardized introduction. A middle school student learning to structure a plan should not delegate this learning to a tool.

This framing is consistent with neuropsychological data. It is also realistic in practical terms: it does not ask teachers to eradicate the use of a tool that students use outside of classes anyway, but rather to build explicit pedagogy around the conditions in which AI enhances rather than replaces.

The Generational Question That No One Can Yet Settle

The real issue is not individual. It is generational and has a specific temporal horizon. The 15-25 year-olds entering higher education and the job market today are the first cohort to have access to generative AI tools during their training phase. If these tools become the dominant mode of writing and structured reasoning for this cohort before basic competencies are solidly established, the effects on cognitive capacities will only be measurable with a delay of ten to fifteen years.

This is exactly the temporality that makes the question politically difficult. The benefits of AI delegation are immediate and visible (time savings, more fluid texts, reduction in felt workload). The cognitive costs, if there are costs, are deferred and diffuse. Political decision-makers and educational institutions rarely work on this horizon. And technology platforms have no incentive to do so.

We can push the reasoning further, provided we hold it as a hypothesis and not a certainty. If writing competency is linked to the capacity for structured reasoning, and if this capacity is partially the product of its regular practice, then an entire cohort that exercised this capacity less during its training phase could present measurable deficits in complex reasoning tasks. The work of Acemoglu and Johnson on technology and the distribution of gains reminds us that the effects of a technology on human capacities depend on who controls its deployment and in what interest. This dynamic is directly relevant here.

This is not prophecy. It is a blind spot that current data does not allow us to fill, but whose logic compels us to name it. The first serious longitudinal study on the cognitive effects of intensive generative AI usage among adolescents does not yet exist. It should.

What Is Already Being Built and What Remains to Be Done

Alarmism would be inaccurate. Actors are building concrete responses, and some are promising. Beyond Khan Academy, research teams in cognitive science are working to define usage protocols that preserve production effort. Educational technology companies are developing AI tools that function in scaffolding mode rather than substitution mode. UNESCO published guidance in 2023 for the use of generative AI in education, precisely insisting on the distinction between uses that enhance and those that bypass.

In Europe, the AI Act imposes transparency obligations on providers of AI tools intended for education regarding their functioning. This is a necessary but insufficient condition: knowing that a tool cognitively delegates does not prevent delegation if the tool’s design is not modified.

The actionable question is one of design. Educational AI tools can be designed to require production effort before offering assistance. This is not a difficult technical question to solve. It is a question of design choice, therefore of regulation or incentives for that choice. A “cognitively preserving” certification for educational AI tools, analogous to nutritional or environmental certifications, is an idea circulating in certain academic circles. It has not yet found a serious institutional sponsor.

What is at stake is not the fate of writing as a cultural practice. It is the question of whether the next generations reach adulthood with structured reasoning capacities formed by effort, or with validation capacities formed by delegation. These two cognitive profiles are not equivalent for what societies must do: maintain institutions, produce new knowledge, exercise judgment in ambiguous situations.

The pharmakon prescribes itself. But someone must be willing to do it.

Sources

  1. Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. https://www.codeandcortex.fr/intelligence-artificielle-anne-alombert/
  2. Anna Alombert — philosophy of technology and pharmakon of generative AI (2025-2026): https://www.codeandcortex.fr/intelligence-artificielle-anne-alombert/
  3. Daron Acemoglu & Simon Johnson, Power and Progress, PublicAffairs, 2023
  4. Khan Academy — Khanmigo, public project documentation (khanacademy.org)
  5. UNESCO, Guidance for generative AI in education and research, 2023 (unesdoc.unesco.org)
  6. Bergounioux-Thiollière Report, School and Digital Technology, 2024 (French Senate)
  7. Kosmyna et al. 2025 study — primary source (arXiv): https://arxiv.org/abs/2506.08872
  8. Official MIT Media Lab page — Kosmyna publication: https://www.media.mit.edu/publications/your-brain-on-chatgpt/
  9. Methodological critique Stanković et al. 2026: https://arxiv.org/abs/2601.00856
  10. Anne Alombert — academic profile and publications: https://organoesis.org/anne-alombert
  11. Daron Acemoglu — MIT Sloan Faculty: https://mitsloan.mit.edu/faculty/directory/daron-acemoglu
  12. Khanmigo — Khan Academy (official site): https://www.khanmigo.ai/
  13. Brookings — AI and educational institutions since 2023: https://www.brookings.edu/articles/should-schools-ban-or-integrate-generative-ai-in-the-classroom/