The Outsourced Brain: The Cognitive Cost of Generative AI

With the widespread and uncritical use of generative AI tools, we are outsourcing critical parts of the cognitive process: synthesis, structuring, and the formulation of ideas. In the short term, this may mean efficiency. In the medium term, it leads to an erosion of knowledge.

Artificial intelligence, when properly prioritised, is indeed fulfilling its initial promise: more productivity, less effort, greater speed. Users of AI tools, particularly generative AI, are on average 60% more productive, as they benefit from immediate, structured, and contextual responses. Work gets done faster and, paradoxically, for longer. For any executive, the benefit and value of AI within the organisation may seem unequivocal.

However, it is not enough to look only at what we gain. We must also understand what we lose.

A recent experiment conducted by MIT, Your Brain on ChatGPT (MIT Media Lab, 2025), compared the performance of three groups of participants in writing an essay: one using an LLM, one using a search engine, and one using no tools. The study reveals an unsettling pattern. As cognitive effort decreases, so too does the depth of learning. When the answer already comes organised, the brain stops doing the essential work: integrating, questioning, and reconstructing. We move from an active mode of thinking to a logic of passive supervision.

The numbers are revealing: 83.3% of LLM users were unable to reproduce any part of their work, compared with only 11.1% in the groups that used search engines or relied solely on their own reasoning. Moreover, not a single participant in the LLM group was able to produce a fully accurate quotation. Zero. By contrast, only 16.7% of participants in the search group and 11.1% in the no-tools group failed to do so.

This differentiated performance across the three groups is not a detail with merely academic implications. It shows that although knowledge was used, it was not assimilated. And that is relevant in any context, whether academic, professional, or otherwise.

The MIT study also tells us that 16% of LLM users reported not feeling that the work belonged to them. This phenomenon did not occur in the other groups. It is not worth pretending that this circumstance has no potentially relevant repercussions for responsibility or accountability in relation to the work produced and the decisions associated with it. In short, and put differently: more is produced, but less is appropriated. Work is executed, but not internalised.

In my view, there is one particular finding that should concern any manager: only 6% of LLM users were dissatisfied with the outcome of their work, compared with 17% in the group that worked without support. In other words, the less one thinks, the more one believes the result is good. Confidence increases precisely when knowledge decreases.

The problem becomes worse when dependency starts early. The data show that those who used AI from the very beginning of the task had poorer memory, a lower capacity for reconstruction, and weaker signs of cognitive integration. By contrast, those who began without support and only later used AI retained more, thought better, and used the tool more strategically. AI amplifies skills. It does not replace them.

With the widespread and uncritical use of generative AI tools, we are outsourcing critical parts of the cognitive process: synthesis, structuring, and the formulation of ideas. In the short term, this may mean efficiency. In the medium term, it leads to an erosion of knowledge.

There is also a silent effect on organisational culture. Less debate, less confrontation of ideas, less collective learning. Interaction with the machine replaces intellectual friction between people. And without friction, there is no critical thinking.

It may seem unusual that I, of all people (someone who leads an organisation that develops and applies AI) should be writing these lines. But the issue I raise here is not technological. It is behavioural and strategic. We may have faster teams, but also more superficial ones, lacking a critical view of the results of their work. We may have more agile decisions, but less well-founded ones. Organisations that produce more, but learn less. And, as we know, an organisation that does not learn becomes dependent on whatever allows it to produce. But we can also have organisations that think more and think better.

The adoption of AI is not the issue, nor is it even an option today. But the way we use it is. If AI is used merely to accelerate, the result is clear: more output, less thought. But if it is designed to preserve cognitive effort, it can become a true competitive advantage.

Ultimately, the choice is simple. We can use AI to think better. Or simply to stop thinking