AI Threatens Cognitive Skills, Study Reveals

A Study Reveals the Impact of AI on Problem Solving

A recent study conducted jointly by researchers from the United States and the United Kingdom highlights a concerning effect of using AI assistants. According to their findings, exposure of just 10 to 15 minutes to an AI used as a response machine can measurably weaken users' problem-solving skills and persistence on subsequent tasks performed without assistance.

The results of this research show that while AI assistance improves immediate performance, it has a notable downside. Once the AI is removed, users who relied on it perform worse than those who approached the same tasks independently from the start. Additionally, they tend to abandon tasks more frequently.

Experiments on Fraction Problems

In the first experiment, participants faced 15 fraction problems, ranging from simple calculations to more complex tasks. One group had access to GPT-5 via a sidebar, where each problem and its solution were preloaded. This allowed participants to obtain correct answers with minimal effort, simply by typing "Answer?". The control group, on the other hand, worked without any assistance tools.

After solving 12 problems, the AI was removed without warning, and all participants had to solve three identical test problems on their own. While the AI was available, the group using the AI nearly solved all the fraction problems. However, once the AI was removed, their success rate dropped below that of the control group, and their abandonment rate significantly increased.

Confirmation by a Second Experiment

A second experiment was conducted to correct a methodological bias observed in the first. In this experiment, lower-performing participants in the AI group could submit correct answers via the AI, preventing them from being filtered by the same criteria as the control group. To address this, a pre-test with simple fraction problems was added, and the control group received a sidebar with pre-test solutions to match the interface of the AI group.

This second experiment replicated the effect with a more rigorous methodology. The AI group again excelled during the learning phase but faltered during the unassisted test. The abandonment rates were roughly equivalent on average between the two groups.

Different Styles of Using AI

The study revealed that 61% of AI users primarily asked for direct answers from the assistant. A quarter of the participants used it to obtain hints or explanations, while the rest did not use it at all. During the pre-test, these groups achieved similar results in terms of resolution and abandonment rates.

However, during the unassisted test, users who asked for "direct answers" performed the worst and abandoned tasks most frequently. In contrast, those who completely ignored the AI displayed the highest success rates. After the AI was removed, the results changed significantly. Individuals who relied on direct answers achieved the worst results, while participants who ignored the AI showed even higher success rates than the control group.

Effect on Reading Comprehension

To verify whether this effect was limited to mathematics, the researchers applied the same experimental design with reading comprehension passages from the U.S. SAT. Here, the control group received a sidebar with general tips to mimic the context change between the learning and testing phases. The team also counted responses given in less than five seconds as abandonments, as the passage cannot be read that quickly.

The results confirmed the mathematical experiments. The AI group provided fewer correct answers during the unassisted test and abandoned significantly more often. The reduction in persistence, according to the researchers, is a general side effect of AI-assisted problem solving, even on tasks closely related to critical thinking.

Two Mechanisms, One Structural Problem

The study proposes two explanations for the observed loss of persistence. First, the AI resets the reference point regarding the difficulty of a task. Working without help then seems more challenging, much like one becomes accustomed to any convenience. This mechanism is self-reinforcing: each shortcut increases the perceived cost of doing the work oneself next time.

Secondly, users miss out on the productive struggle that builds both knowledge and a realistic perception of their own abilities. The researchers link their findings to the broader debate on the gradual loss of skills. AI systems optimized for instant assistance could undermine their users' long-term capabilities.

Fractions and reading comprehension may seem easy to delegate, but they are prerequisites for more advanced skills like algebra and critical thinking. Students with fewer academic resources are particularly at risk. If just 10 minutes of use produces measurable effects, the researchers warn that the consequences could accumulate over months and years, becoming difficult to reverse.

A Growing Body of Evidence on the Cognitive Costs of AI

Previous research has also pointed in the same direction, albeit through less rigorous methods. A study from the Swiss Business School found a strong negative correlation between AI use and critical thinking, particularly pronounced among participants aged 17 to 25. Higher education acted as a protective factor: individuals with more schooling questioned AI-generated information more often and engaged in deeper analysis.

A joint study by Microsoft Research and Carnegie Mellon described an "irony of automation": by managing routine work, AI tools deprive users of the chance to exercise their "cognitive muscles." For routine or low-stakes tasks, users simply turn to AI.

An Anthropic study with 52 software developers, primarily juniors, also showed that AI assistance can hinder the learning of new programming skills. Participants were asked to solve two tasks using the unfamiliar Trio library. One group had access to a GPT-4o-based assistant; the control group worked solely with documentation and web research.

In a follow-up knowledge test, the AI group scored 17% lower. Again, how people used the tool mattered: those who asked for explanations learned better than those who delegated the work.

User experience also counts. In another Anthropic study, experienced users of Claude achieved success rates about four percentage points higher than new users on identical tasks. They worked iteratively with the model rather than simply issuing commands.

Other research shows that AI can enhance individual and team performance. However, many companies still struggle to translate these isolated productivity gains into real improvements in efficiency or revenue growth for various reasons.