You may be quitting sooner because of your AI.
Not because you are lazy. In three experiments with 1 222 people, those who had an AI assistant for just 10 to 15 minutes solved fewer problems on their own once it was gone, and skipped more of them. The people who asked it for straight answers did worst.
Here is what happened, and how to use AI without paying that price.
The AI group was ahead until the AI disappeared
Researchers from Carnegie Mellon, UC Berkeley, Oxford, MIT and UCLA recruited people online in the US and gave them problems: fractions in two experiments, SAT-style reading questions in the third.
Half had an AI assistant (GPT-5) in a sidebar. It knew the answers, so typing “answer?” got you one. The other half worked alone. Then, without warning, the AI vanished, and everyone did the same final problems unaided.
Wrong answers carried no penalty, and anyone could hit “skip”. So skipping wasn’t about ability. It was a choice not to try.
With AI, people solved more and skipped less. Without it, the advantage flipped.
Without the AI, people solved less and skipped nearly twice as often
In the first fractions experiment, on the final problems:
- 57% vs 74% solved, the AI group against the solo group. That’s roughly a quarter fewer right answers.
- 20% vs 11% skipped. Nearly twice as many gave up.
Reading told the same story:
- 76% vs 89% solved.
- 8% vs 1% skipped.
A second, larger fractions experiment found the same drop in accuracy. The overall rise in skipping was smaller there and not statistically clear. So the researchers looked at how people had used the AI.
Asking for answers did the most damage
In that experiment, people reported how they had used the assistant. About 6 in 10 had mostly asked it for straight answers.
- 65% solved, 13% skipped: people who asked for answers.
- 76% solved, 5% skipped: people who asked for hints or explanations.
- 89% solved: people who had the AI but didn’t use it.
The answer-seekers ended up below their own starting level. The hint-seekers held steady or improved.
People chose their own strategies here, so this comparison doesn’t prove cause. But the pattern is hard to ignore. Getting the answer handed to you is what seems to hurt.
If you ask for the answer, you practise asking. If you ask for a hint, you still practise solving.
Instant answers make your own effort feel slow
The researchers suggest two reasons.
First, your sense of how long things should take shifts. When answers arrive in seconds, working it out yourself starts to feel painfully slow. It’s like fast Wi-Fi: once you’ve had it, normal speed feels unbearable.
Second, you never find out what you can do. Struggling through a problem teaches you the answer, and also that you are able to find it. Skip the struggle and you miss that second lesson.
On your own, giving up then looks like the sensible choice.
Persistence is a skill you build by struggling
Learning researchers call it desirable difficulty. Testing yourself, or wrestling with a problem before seeing the solution, feels harder in the moment and makes learning stick better.
AI is brilliant at removing that discomfort.
That’s great when you just need the job done. It’s a problem when you are trying to get better at something.
You can use AI and still do the hard part
These are reasonable inferences from the findings, not things the study tested:
- Try first. Give a hard problem a real go before you open the chatbot.
- Ask for hints, not answers. “What’s the first step?” beats “What’s the answer?”
- Use AI to check, not to start. Solve it, then compare.
- Know your goal. If you just need it done, use AI freely. If you are learning, do the hard part yourself.
What the study can’t tell you
- Sessions were short, about 10 to 15 minutes. Whether the effect builds up or fades with daily use is unknown.
- Testing was immediate. Nobody knows yet whether the drop lasts hours or days.
- The tasks were simple and lab-style: fractions and reading questions done by online volunteers, not students in class or people at work.
- The AI was maximally helpful and gave full answers on demand. A tutor-style AI might work differently.
- Not every result was clear-cut. In one experiment the overall rise in skipping wasn’t statistically significant.
- The paper was presented at COLM 2026, a peer-reviewed AI conference.
Try before you ask
AI can do the hard part for you. Only you can keep the skill.
Next time you’re stuck, give the problem a few honest minutes, then ask for a hint instead of the answer.