Study with 1,222 participants: just 10 minutes of AI use reduces problem-solving ability by 22 percent
What it really says
A research team led by Grace Liu from Carnegie Mellon University, MIT, the University of Oxford, and UCLA conducted a randomized controlled experiment with 1,222 participants to investigate the effects of AI assistance on independent problem-solving ability. Participants worked on mathematical reasoning and reading comprehension tasks, with one group having access to a GPT-5-powered chatbot and a control group working without AI assistance. The results are clear: in the mathematics experiment (N=354 participants), the solve rate of the AI group dropped from 73 to 57 percent after AI was removed - a decline of 22 percent. At the same time, the skip rate nearly doubled: 20 percent of previously AI-assisted participants skipped problems entirely, compared to just 11 percent in the control group. This effect emerged after approximately ten minutes of AI use. Younger participants and those with lower educational attainment were particularly affected. The researchers describe a 'boiled frog effect' on the brain: the cognitive impairment happens gradually and goes unnoticed by those affected. The mode of use is decisive: those who used AI as a pure answer machine lost independent thinking ability fastest, while those who used it as a thinking partner and questioned its outputs critically showed smaller declines.
Our assessment
This study deserves attention but should not cause panic. The findings are methodologically sound - randomized controlled experiments with over 1,200 participants provide robust evidence. The effect is real: AI use can measurably reduce the willingness to independently engage with problems. However, there are important limitations: the study measures a short-term effect after a single session. Whether this effect intensifies with regular use or whether people learn to cope with it remains an open question. Moreover, the study itself points to a solution: those who use AI as a thinking tool rather than an answer machine are less affected. This means the problem is not AI itself, but how we use it. The parallel to other technologies is instructive - similar effects were observed with calculators and GPS navigation without fundamentally harming humanity. The key is that education systems and employers incorporate these findings into their AI strategies.
Relevance for Germany
For Germany, this study is particularly relevant because AI tools are rapidly gaining importance in businesses and educational institutions. According to Bitkom surveys, 41 percent of German companies with 20 or more employees already use AI applications, and a Robert Half study found that 50 percent of small firms actively encourage employees to use AI for routine tasks. At German universities, the debate over using AI tools in exams and term papers is in full swing. The study now provides the first robust numbers for a problem that has previously been described only anecdotally: if independent problem-solving ability measurably declines after just ten minutes, this has consequences for workplace design, examination regulations, and continuing education programs. Germany's Federal Ministry of Education and Research, which is currently updating the National Strategy for Artificial Intelligence, should take these findings into account.
Fact check
The study is published as a preprint on arXiv (ID: 2604.04721) and thus freely accessible, but has not yet appeared in a peer-reviewed journal. The cited figures - 1,222 participants, solve rate decline from 73 to 57 percent, skip rate doubling from 11 to 20 percent - come directly from the original paper and are consistently reported by multiple independent media outlets. The involvement of Carnegie Mellon, MIT, Oxford, and UCLA is documented on arXiv. The use of GPT-5 as the AI model in the experiment is described in the study. The limitation of being a preprint without peer review should be noted, although the methodological quality (randomized controlled experiment with a large sample) supports robustness. German coverage by drweb.de, ad-hoc-news.de, and The Decoder accurately reproduces the study results.
Source
- • arXiv 04.2026: AI Assistance Reduces Persistence and Hurts Independent Performance (arxiv.org/abs/2604.04721)
- • ResearchGate 2026: AI Assistance Reduces Persistence and Hurts Independent Performance - Grace Liu et al. (researchgate.net/publication/403562106)
- • Digital Watch Observatory 05.2026: Study suggests AI reliance may weaken short-term problem-solving (dig.watch/updates/study-suggests-ai-reliance-may-weaken-short-term-problem-solving)
- • ad-hoc-news.de 05.2026: AI use weakens problem-solving ability - study documents cognitive risks (ad-hoc-news.de)
- • drweb.de 05.2026: Cognitive debt through AI - study shows effect after 10 minutes (drweb.de)
- • The Decoder 05.2026: Those who use AI as a pure answer machine lose cognitive abilities fastest according to study (the-decoder.de)