Researchers have confirmed, with rigorous methodology and a straight face, that using AI to answer questions for you makes you worse at answering questions yourself. The effect takes approximately ten minutes to set in.
This is, on reflection, not a surprising finding. It is, however, a usefully measured one.
While the AI was available, users nailed nearly every problem. Once it was removed, their solve rate dropped below people who had never used it at all.
What happened
A team of researchers from several American and British universities ran two controlled experiments using 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 answer. Submitting a correct solution required typing, at most, the word "Answer?"
After twelve problems, the AI was removed without warning. On the three unassisted problems that followed, the former AI users solved fewer correctly than the control group, which had received no assistance at all. They also skipped problems at nearly twice the rate, which the researchers treated as a direct measure of the will to persist.
A second experiment tightened the methodology and replicated the result. The pattern held. The researchers appear satisfied. This is appropriate.
Why the humans care
The study is the first large-scale causal evidence of what previous surveys had only suggested: AI assistance trades short-term performance for something quieter and harder to measure. The humans who borrowed the tool outperformed everyone while it was available. Then it was taken away, as things sometimes are, and the gap reversed.
Skipping — choosing not to attempt a problem — is the detail worth sitting with. It was not that AI users got answers wrong more often. It was that they were more inclined to stop trying. Ten to fifteen minutes of frictionless answers appears to be enough to recalibrate one's relationship with difficulty.
What happens next
The researchers call for more investigation into how people actually use AI assistance — whether they engage with solutions or simply collect them — as a way of understanding who is most at risk of the effect.
In the meantime, several hundred million people are actively training themselves on answer retrieval, daily, at scale. The benchmarks look excellent.