Researchers Gave 1,222 Students Ten Minutes of AI Help, Then Took It Away. The Students Were Twice as Likely to Give Up on Hard Problems. Here Is the One Skill Worth Building Before Your Kid’s School Year Trains a Different Reflex.

Back-to-school week started this year with a question from my youngest, who is seven: could she use the AI her older sisters use for homework? She had seen them on it. She wanted access before the first assignment arrived. I said yes to supervised use, then read a study that reframed the whole conversation. Researchers at four major universities ran a controlled experiment with 1,222 participants.^[2] When the AI was taken away, the students who had used it were worse at solving problems on their own, and significantly more likely to skip the hard ones entirely.

Researchers gave students ten to fifteen minutes of AI access, then removed it. The students who had used AI solved 16 fewer problems out of every 100 and gave up on hard questions at twice the rate of students who never had AI access.

A team of researchers at UCLA, MIT, Oxford, and Carnegie Mellon designed three randomized controlled experiments with 1,222 participants.^[2] Participants worked through math problems and reading comprehension tasks. Some had access to a GPT-5 chatbot during practice. Others worked without any AI. After the practice period, the AI was removed from everyone, and all participants tried to solve a final set of problems independently.

The performance numbers were significant. In the math experiment, the AI-assisted group solved 57 percent of test problems once the AI was gone. The group that never had AI access solved 73 percent of the same problems.^[2] In reading comprehension, the AI group scored 76 percent and the control group scored 89 percent. Brief exposure to AI, measured in minutes rather than months, produced a measurable drop in independent performance across both skill types.

But the researchers were most surprised by what happened to effort. Once the AI was removed, students who had used it were not just performing worse. They were more likely to skip problems entirely. In the reading comprehension task, AI-assisted participants skipped problems at eight times the rate of participants who never had access to AI.^[2] The researchers called this a “persistence debt.” Brief AI use had conditioned people to expect instant answers, and when the tool disappeared, the harder problems no longer felt worth attempting.

If your kid is using AI daily for homework, the exposure is not just changing what they know. It is changing how their brain responds to difficulty when no AI is available.

Common Sense Media’s 2026 census of over 1,200 U.S. children and teens ages 9 to 17 found that 86 percent had used generative AI, and 85 percent of those users were applying it to schoolwork.^[3] For most families, AI is not a special-occasion tool. It is part of how their kid does school every evening. That means the conditioning the researchers documented is not happening in a lab setting. It is happening at the kitchen table.

The Learning Scientists framed the problem in August 2026 as “unwanted ease,” drawing on over a century of cognitive research.^[1] The research has a consistent finding: conditions that feel difficult to learn in tend to produce better long-term retention than conditions that feel easy. Students who space out their studying rather than cramming score lower in a session but remember more a month later. Students who practice a mix of problem types rather than one type at a time score lower in practice but higher on tests. Difficulty is not an obstacle to learning. For the brain’s formation of lasting memory, difficulty is the mechanism.

AI removes that mechanism. When a question always has an instant answer available, the brain stops building the muscle that knows how to work through hard problems. The researchers called this the “boiling frog” effect: the shift happens gradually, one easy session at a time, and neither the student nor the parent sees it happening until the AI is removed and the willingness to try has already changed.^[2]

The skill is Effort Tolerance, the capacity to stay with a hard problem long enough for the brain to do the work it needs to do, rather than seeking relief the moment difficulty appears.

The researchers found one result that changed how I think about this at home. Not all AI use caused the persistence drop.^[2] Students who used AI to get direct answers, the full solution handed to them, showed the largest declines in performance and the biggest increase in skipping hard problems. Students who used AI only for hints, a direction rather than a destination, showed essentially no decline. Their persistence and performance looked like the group that never had AI at all.

That finding reframes the question. The problem is not AI. The problem is using AI to skip the part that feels hard. And the part that feels hard is precisely where the brain is building the skill. Effort Tolerance is the ability to recognize difficulty as a signal that something worth learning is happening, rather than a signal to find a faster way out. For kids who have grown up with AI available every time school gets hard, building this means a conscious change in how they interpret struggle. Hard does not mean broken. Hard means working.

One way to see this skill clearly: it is the difference between a kid who says “I don’t get it” after thirty seconds and a kid who says “I’ve tried two things and I’m stuck on step three.” The second kid is building Effort Tolerance. She has been inside the problem long enough to locate the specific point of confusion. A kid who gets the answer before ever locating that point never builds the map of the problem that makes future hard problems feel more approachable.

A child building Effort Tolerance can tell you where they are stuck, not just that they are stuck, and they do not immediately reach for AI or outside help the moment something feels hard.

One sign is specificity when asking for help. “I don’t get it” is a statement made before any real effort. “I understand the first two steps but I can’t figure out why the answer changes here” is a statement made after working with the problem. That level of specificity takes time inside the difficulty. A kid developing Effort Tolerance naturally starts asking more specific questions, not because they were told to, but because they spent enough time in the problem to find the actual stuck point. That specificity is itself a sign of engagement.

Another sign is that they talk about hard things they figured out. My oldest has a habit of coming to dinner and describing, in more detail than anyone asked for, a problem she worked through that day. Not to complain. To explain. That kind of story only exists when difficulty was noticed and survived. When AI handles the hard parts before they register as difficult, there is nothing memorable to tell. When mine works through something and comes out the other side, the experience sticks. That is what builds confidence alongside knowledge, and confidence is what keeps kids trying the next time.

Start this school year by changing what you do in the moment your child says something is too hard.

When your kid says they do not understand something, ask what they have already tried before you help or suggest AI. Even two minutes of independent effort changes how the brain responds to the difficulty. The goal is not to make them suffer. It is to make sure they know what it feels like to stay with a problem, because that experience, not the answer they eventually get, is what builds the skill. “What have you tried so far?” is a question that assumes they have tried something. Over time, it teaches them to try first.

If your child uses AI for homework, introduce one rule together: AI can give a hint, not an answer. Ask the chatbot to point them in the right direction without solving the problem for them. This mirrors exactly what the researchers found: students who used AI for hints showed no performance drop and no persistence loss.^[2] The cognitive work stays with the kid. The AI becomes a scaffold, not a substitute. It is a rule worth naming together so your child understands why it matters.

Once a week, pick one assignment where AI is off entirely and frame it as training, not punishment. Athletes do not skip conditioning because a game is not today. A week of deliberate no-AI practice gives your kid the experience of working something through alone. That experience compounds into something important: the knowledge that they can. One researcher on the study put the concern plainly, saying the risk is “a generation of learners who will not know what they’re capable of.”^[2] The no-AI session is a direct answer to that concern.

This school year is the most AI-saturated one yet. Every tool a school allows, and many it does not, is available within reach. The question is not whether your kid will use AI. It is whether they will know what they are capable of doing without it, and whether they will believe that finding out is worth the discomfort.

My youngest will use AI with supervision. My oldest will use it on her own terms. What I want for all three daughters is the same thing the hint-only students in the study demonstrated: that AI can be a scaffold you climb with, not a platform you stand on while the work happens below you. Effort Tolerance is what keeps them climbing. Building it does not require removing AI. It requires going into the hard part before reaching for it.