Just ten minutes of using AI as an answer machine can measurably erode problem-solving skills, new study finds Just 10 to 15 minutes with an AI assistant is enough to measurably weaken problem-solving ability and persistence on later tasks done without AI, according to a new study from researchers in the US and UK. The research, conducted by teams at several American and British universities, shows that while AI assistance boosts immediate performance, it comes with a catch: once the AI is taken away, users perform worse than people who tackled the same tasks on their own from the start. They also give up more often. Previous evidence of these effects has mostly come from surveys or small samples, the researchers note.

This is the first large-scale causal evidence drawn from controlled experiments. Fraction problems expose the crutch effect In the first experiment, participants worked through 15 fraction problems ranging from simple one-step calculations to more complex three-step tasks. One group had GPT-5 available in a sidebar, preloaded with each problem and its solution. That meant participants could get correct answers with almost no effort. Just typing "Answer?" was enough. The control group worked with no tools at all. After 12 problems, the AI was removed without warning, and everyone solved three identical test problems on their own. On those test problems, the former AI users got significantly fewer correct answers than the control group. They also skipped problems almost twice as often.

Since there was no penalty for wrong answers and pay wasn't tied to performance, the researchers treated skipping as a direct measure of persistence and motivation. A second experiment confirms the pattern A follow-up experiment addressed a methodological flaw: in the first run, weaker participants in the AI group could submit correct answers through the AI, meaning they weren't filtered out by the same criteria as the control group. This time, a pre-test with simple fraction problems was added, and the control group got a sidebar with pre-test solutions to match the AI group's interface. The results held up: the AI group again underperformed the control group on the unassisted test. A higher skip rate pointed the same way, though it didn't reach statistical significance overall.

The researchers point to differences in how people actually used the AI as a possible reason. Direct-answer users pay the biggest price About 61 percent of AI users said they mainly asked the assistant for direct answers. Another quarter used it for hints or explanations, and the rest didn't use it at all. On the pre-test, these groups performed the same on both solve and skip rates. Baseline ability and motivation were comparable. After the AI was pulled, the results shifted sharply. People who had relied on direct answers did the worst, while participants who ignored the AI altogether posted the highest solve rates, even higher than the control group. The direct-answer users also declined relative to their own pre-test scores, while the other groups held steady or improved.

The data suggests the negative effects are concentrated in users who outsource the thinking. The same pattern shows up in reading comprehension To see whether the effect was limited to math, the researchers ran the same design with reading comprehension passages from the US SAT. Here the control group got a sidebar with general test tips to mirror the context switch between learning and test phases. The team also counted answers given in under five seconds as skips, since the passage can't be read that fast. The results matched the math experiments. The AI group got fewer questions right on the unassisted test and skipped significantly more.

Reduced persistence, the researchers say, is a broad side effect of AI-assisted problem solving, even on tasks tied closely to critical thinking. Two mechanisms, one structural problem The study points to two explanations for the loss of persistence. First, AI resets the reference point for how hard a task should feel. Working without help then feels tougher, the same way you get used to any convenience. The mechanism is self-reinforcing: every shortcut raises the perceived cost of doing the work yourself next time. Second, users miss out on the productive struggle that builds both knowledge and a realistic sense of their own abilities. The researchers tie their findings to the broader debate about gradual skill loss.

AI systems optimized for instant helpfulness could undermine their users' long-term abilities. Fractions and reading comprehension might look like easy things to delegate, they note, but they're prerequisites for higher skills like algebra and critical thinking. Students with fewer academic resources are especially at risk. If just 10 minutes of use produces measurable effects, the researchers warn, the consequences could compound over months and years and become hard to reverse. User-side fixes like Socratic AI or usage limits are just "band-aids," they argue.

What's needed is a rethink of how these systems are built, away from short-term user satisfaction and toward designs that foster autonomy and sometimes withhold help. A growing body of evidence on AI's cognitive costs Earlier research has pointed in the same direction, though with weaker methods. A study by the Swiss Business School found a strong negative correlation between AI use and critical thinking, most pronounced among participants aged 17 to 25.

Higher education acted as a protective factor: people 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 handling routine work, AI tools deny users the chance to flex their "cognitive muscles." For routine or low-stakes tasks, users simply default to the AI. An Anthropic study with 52 mostly junior software developers similarly showed that AI assistance can hurt 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 only with documentation and web search. On a follow-up knowledge test, the AI group scored 17 percent lower. Again, how people used the tool mattered: those who asked for explanations learned better than those who offloaded the work. Usage experience matters too. In another Anthropic study, experienced Claude users hit success rates about four percentage points higher than newcomers on identical tasks. They worked iteratively with the model rather than just issuing commands. Other research shows AI can boost individual and team performance.

But many companies still struggle to turn those isolated productivity gains into real efficiency improvements or revenue growth, for plenty of reasons. AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe nowAI news without the hype Curated by humans. - More than 16% discount. - Read without distractions – no Google ads. - Access to comments and community discussions. - Weekly AI newsletter. - 6 times a year: “AI Radar” – deep dives on key AI topics. - Up to 25 % off on KI Pro online events. - Access to our full ten-year archive. - Get the latest AI news from The Decoder..