MIT Media Lab (2025): "Cognitive Debt"
The protocol. Nataliya Kosmyna and her colleagues followed 54 participants for four months, divided into three groups to write essays: with ChatGPT, with a search engine, or without any tools. Brain activity was measured using electroencephalography. During a fourth session, some participants switched conditions.
The results. The tool-free group showed the strongest and most extensive brain networks, while the ChatGPT group showed the weakest. ChatGPT users struggled to recall their own text and expressed the lowest sense of authorship. Those who switched from ChatGPT to writing alone showed reduced engagement; conversely, those who first wrote alone and then used ChatGPT showed better recall.
The limitations. This is a preprint based on a small sample, which has received critical comments. Mustapha Benkalfate cites it because other studies converge on similar findings, and because its practical lesson is robust: start alone, then use AI to improve, rather than the other way around.
The Lancet Gastroenterology & Hepatology (2025): Deskilling of Physicians
A multicenter observational study compared colonoscopies performed without AI before and after the introduction of a detection assistance tool. The endoscopists' adenoma detection rate, without assistance, dropped from 28.4% to 22.4%. This is an example of what Mustapha Benkalfate calls rust: an acquired skill that erodes when delegated.
Microsoft Research and Carnegie Mellon (CHI 2025): Trust Versus Critical Thinking
A survey of 319 knowledge workers, covering 936 real-world use cases, shows that the higher the trust in AI, the less critical thinking effort people report; conversely, confidence in one's own skills is associated with more critical thinking. The work shifts from execution to supervision. This is the "worst combination" described by Mustapha Benkalfate: rising trust while skills erode.
British Journal of Educational Technology: Metacognitive Laziness
Yizhou Fan and his colleagues compared several forms of learning assistance, including ChatGPT. AI improved performance on the immediate task, without an equivalent gain in knowledge or transfer. The authors speak of "metacognitive laziness": delegating the regulation of one's own learning. Hence Mustapha Benkalfate's emphasis on metacognition.
The Ironies of Automation, from 1983 to Generative AI
In 1983, Lisanne Bainbridge showed in Automatica that automation leaves humans with the most difficult tasks, while eroding the skills they need to perform them. In 2024, Auste Simkute, Lev Tankelevitch, Abigail Sellen, and their colleagues from Microsoft Research and University College London applied this framework to generative AI: work shifts from production to evaluation, workflows are reorganized, interruptions multiply, and easy tasks become easier while difficult tasks become harder.
What Mustapha Benkalfate Takes Away for Organizations
- The danger is not the tool, but the substitution. Using AI to improve one's own work leads to progress; using it to replace it leads to regression.
- Deskilling is difficult to reverse. Once a skill is lost, regaining it requires deliberate effort.
- A "cognitive gym" is needed. After the car and industrial food, we invented gyms; after social media and the lazy use of AI, we need to organize voluntary friction for high-value tasks.
- Human risks must be part of governance. Because they are difficult to quantify, leaders often dismiss them in favor of measurable gains. He suggests mapping them job by job and tracking them over time.
The definition of each concept, from cognitive debt to automation bias, is in our glossary of cognitive risks of AI.