Cognitive debt: when AI makes the work easier and the worker weaker

The most expensive thing generative AI does to a knowledge worker is also the hardest to put on a dashboard. It saves them time. It also, on the current evidence, risks hollowing out the judgement that time was meant to build. Both things are true at once, and the gap between them is where the management problem lives.

Start at the unsettling end of the literature. In a controlled experiment, 54 people wrote essays under one of three conditions, unaided, with a search engine, or with ChatGPT, while EEG sensors tracked their brain activity. The group using the chatbot showed the weakest and least distributed neural engagement of the three, and the deficit widened the longer the study ran. The researchers called the effect 'cognitive debt'.¹ A survey of 319 knowledge workers points the same way from a different angle: the more confidence people had in generative AI, the less critical thinking they reported doing, as their effort moved from producing work to checking the machine's version of it.² A separate analysis found frequent AI use associated with lower critical-thinking scores, with cognitive offloading as the mechanism that carries the cost.³

This is not the claim that AI makes people stupid, and anyone making that claim is overreaching. The same tools raise output in ways that are well measured and large.

The other ledger

A randomised trial of 453 professionals found ChatGPT cut task-completion time by around 40 per cent and lifted quality by about 18 per cent, with the biggest gains going to the weakest performers. A separate customer-support deployment showed a 15 per cent average productivity gain, rising to 36 per cent for workers in the bottom skill quintile.⁴ Read on their own, these are the numbers of a levelling technology: it transmits expert practice to novices and narrows the gap between the best and the rest. An honest account has to keep that evidence in view. Corrosion of capability isn't the charge here. The point is narrower, and more awkward: a gain in output is not a gain in underlying skill, and the two can drift apart without anyone noticing.

The place that distinction bites hardest is the bottom of the career ladder. AI helps novices most, yet firms adopting it have cut hiring of 22-to-25-year-olds by roughly 13 per cent in exposed occupations.⁵ The cohort that gains the most from the tool is the cohort being hired the least. That thins the population of juniors doing the foundational, repetitive, slightly inefficient work through which expertise is actually built. What looks like a productivity win is also a leadership-pipeline problem on a delay: the firm books the efficiency this quarter and pays for the missing senior judgement several years later, by which point the cause is invisible.

Why the dashboards cannot see it

The metrics most organisations watch count AI adoption, and more adoption reads as more value. A tool-usage dashboard cannot tell the difference between a team using AI at the top of its licence and a team quietly becoming dependent on it.¹² The shift from producing to verifying looks, on the numbers, identical to healthy productivity. The danger is not that the work gets worse tomorrow; it is that the capacity to do the work without the tool wastes away while output holds flat, until something genuinely hard arrives and the shallow grasp gives. The literature has a name for the intermediate state, an illusion of competence, in which fluent output masks an understanding that was never built.⁶

Seeing it does not mean measuring 'critical thinking', which no instrument can do directly, and it does not mean surveilling individuals, which corrupts the signal it claims to read. The observable thing is the behavioural shift: whether, across a team and over time, work is drifting from production toward passive verification. Aggregated, consent-based workstyle signals can surface that drift without ranking a single named employee, which is the only version of this measurement that does not poison its own data.

The prescription follows from the diagnosis, and restriction is not it. Capture the upside, which is real, while watching the two places the downside gathers: where checking has replaced doing, and where juniors are being spared the practice that would have made them senior. A healthy division of labour keeps people producing, not only approving. So the question for any leader who has just celebrated an AI productivity number is uncomfortably simple. Do you know whether your people are getting better, or only getting faster?

Footnotes

1. Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv:2506.08872. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt ↩ ↩²

2. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Microsoft Research. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/ ↩ ↩²

3. Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. (Summary via Psychology Today, 2025.) https://www.psychologytoday.com/us/blog/the-art-of-critical-thinking/202512/is-generative-ai-rewiring-our-brains-heres-how-it-happens ↩

4. International Center for Law & Economics. (2025). AI, productivity, and labor markets: A review of the empirical evidence. https://laweconcenter.org/resources/ai-productivity-and-labor-markets-a-review-of-the-empirical-evidence ↩

5. Filippucci, F., et al. (2025). AI and jobs: A review of theory, estimates, and evidence. arXiv:2509.15265. https://arxiv.org/html/2509.15265v1 ↩

6. International Journal of Research and Scientific Innovation. (2025). Illusion of competence and skill degradation in artificial intelligence dependency among users. IJRSI. https://rsisinternational.org/journals/ijrsi/articles/illusion-of-competence-and-skill-degradation-in-artificial-intelligence-dependency-among-users ↩