Oxford/MIT study: Just 10 minutes of AI use measurably reduces problem-solving ability
What it really says
A study by researchers from Oxford, MIT, UCLA, and Carnegie Mellon involving 1,222 participants shows that just 10 to 15 minutes of AI-assisted problem-solving measurably reduces independent performance. In three randomized controlled experiments, participants solved math and reading comprehension tasks - one group with access to a GPT-5-based chatbot, the other without. Results: when AI access was removed, the formerly AI-assisted group solved only 57 percent of math problems correctly, compared to 73 percent for those who never used AI. For reading comprehension: 76 percent versus 89 percent. Additionally, former AI users abandoned tasks more frequently and showed less persistence. The researchers call this the 'boiling frog' effect: each individual instance of AI use feels harmless, but the cumulative impact on cognitive effort becomes difficult to reverse over time. Important differentiation: the effect primarily occurred in participants who used the chatbot to generate complete answers. Those who used AI only for hints or clarification showed no significant decline.
Our assessment
This study provides the first experimental evidence for a fear many people have intuitively: that excessive AI use may atrophy independent thinking. The concern is partially justified, but with important caveats. First: the negative effect occurs primarily when AI is used as an 'answer machine' - those using it as a thinking tool are barely affected. This is a crucial nuance lost in many headlines. Second: the study is a preprint and has not yet been peer-reviewed. Third: 10-15 minutes is a short exposure time - the long-term effect remains unclear. Fourth: participants were predominantly students; whether results transfer to professionals is unknown. Nevertheless, the finding offers an important signal: how we use AI determines whether it makes us smarter or more dependent. Educational institutions and employers should focus on 'AI competency' rather than mere 'AI usage.'
Relevance for Germany
The study is particularly relevant for Germany because AI tools are currently being rolled out across schools and universities. The Standing Conference of Education Ministers adopted guidelines for AI in education in 2025 that explicitly promote chatbot use. The study shows that the mode of integration is decisive: AI as a thinking aid is unproblematic; AI as a substitute for independent thought is harmful. German universities currently developing AI policies receive an important recommendation here. This is also relevant for corporate training: if employees primarily use AI to delegate tasks rather than understand them, problem-solving competency in organizations could decline long-term. The study supports the position of German education experts who advocate for reflective, competency-building AI use.
Fact check
The study is available as a preprint on arXiv (arXiv:2604.04721v2) and has been consistently reported by multiple media outlets. Authors are from Oxford, MIT, UCLA, and Carnegie Mellon - all reputable research institutions. Specific numbers (57% vs. 73% for math, 76% vs. 89% for reading comprehension, 1,222 participants, three experiments) are consistently reported. Important caveat: the study is a preprint and has not undergone peer review. Media coverage tends to omit the differentiation (effect primarily occurs with 'answer delegation,' not with 'hint usage'). The chatbot used was based on GPT-5 - whether results transfer to other AI systems remains open.
Source
- • Liu, G., Christian, B., Dumbalska, T., Bakker, M. A. & Dubey, R. (2026): AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv:2604.04721v2
- • Futurism 04.2026: AI Use Appears to Have a 'Boiling Frog' Effect on Human Cognition, New Study Warns (futurism.com/artificial-intelligence/ai-boiling-frog-human-cognition-study)
- • MobileSyrup 08.05.2026: Study shows 'heavy cognitive cost' of relying on AI for even 10 minutes (mobilesyrup.com/2026/05/08/ai-use-negative-brain-impact-study/)
- • Euronews 16.04.2026: Using AI for basic tasks damages a person's intellect in just 10 minutes (euronews.com/next/2026/04/16/using-ai-for-basic-tasks-damages-a-persons-intellect-in-just-10-minutes-study-shows)
- • The Slow AI Newsletter 04.2026: Does AI Make You Worse at Learning? A New Study of 1,222 People Says Yes (theslowai.substack.com)