AI Doesn’t Just Make You Worse. It Makes You Stop Trying.

A new study finds measurable deskilling after ten minutes of AI use. The people asking it for answers are the ones it breaks first.

Ten minutes of AI use is enough to start eroding your persistence.

That is the finding of a new preprint from Carnegie Mellon, Oxford, MIT and UCLA (With thanks to Paul Austin-Menear for alerting me to this research in the first place).

They ran three randomised controlled trials with 1,222 participants across mathematical reasoning and reading comprehension. The results are consistent, causal, and brief to establish.

You use AI for ten to fifteen minutes. It answers your questions. You feel helped. When the AI is taken away, you perform worse on the same type of problem than people who never used it, and you are more likely to give up. The effect shows up in the data within the first session.

In this post I will:

• Walk through what the study actually found and why the methodology matters.

• Isolate the finding that changes the conversation: the people who used AI for direct answers showed the largest persistence drop. The people who used it for hints did not.

• Give paid subscribers a practical three-part framework for using AI so the scaffolding survives.

What the study found

Previous evidence on AI-induced deskilling has been correlational or from small samples. This paper is the first large-scale causal evidence.

The setup was simple. Participants were given fraction problems of increasing difficulty. Half received an AI assistant (GPT-5) in a sidebar. Half did not. Both were allowed to skip problems with no penalty. Skipping, not failing, was the persistence measure. After a main phase, the AI was removed and both groups solved the same final problems unassisted.

The findings, replicated across three experiments:

1. With AI available, AI-assisted participants performed better and skipped less. This is predictable.

2. Without AI, AI-assisted participants performed worse and skipped more than the control group. On the final three unassisted problems in Experiment 2, the AI group had a solve rate of 71% against 77% for controls. Skip rate was higher too. The effect replicated in reading comprehension with a larger effect size as well.

3. The effect emerged in ten to fifteen minutes. Participants were not AI veterans. They were not in a long-term dependency spiral. Ten minutes of assisted work was enough to measurably change how they behaved when the AI was gone.

Short-term help, measurable medium-term cost.

---

The finding that reframes the conversation

Buried in Experiment 2 is the finding that should change how we talk about AI and learning.

The researchers asked participants in the AI group how they had used the assistant. Three groups emerged:

• 61% used the AI to get direct answers.

• 27% used it for hints or clarifications.

• 12% said they did not use it at all.

At pretest, these three groups performed identically. There was no pre-existing skill difference that explained their later performance gaps. Random assignment held.

At post-test, with AI removed, the groups had diverged sharply. The direct-answer group had the lowest solve rate (65%), the highest skip rate (13%), and showed a decline of 10% in solve rate from their own pretest baseline. The hints group and the non-users performed the same as or better than control.

This is the crux. AI did not hurt everyone who used it. It hurt the people who used it to skip the thinking. The people who used it to scaffold their own thinking were fine.

The paper puts it this way:

“When people prompted AI to solve tasks for them, they were less likely to persist with tasks compared to people who didn’t use AI or people who used AI to aid understanding.”

This is the same distinction that every Slow AI post on critical AI literacy has drawn. It is the line between using AI to help you think and using AI to replace your thinking. The paper has now put a number on the cost of the latter.

---

Why persistence matters more than performance

The usual story about AI and deskilling focuses on capability. Students who use AI will not know how to write, reason, or do maths without it. That is the familiar worry.

This study says the capability loss is real but the smaller part of the problem. The bigger loss is motivational.

Persistence, the disposition to keep trying when something is hard, is one of the strongest predictors of long-term learning. It predicts academic achievement, workforce adaptability, resilience, and the ability to acquire new skills later in life. It is what keeps a student working through the tenth algebra problem when the first nine felt impossible.

The paper argues, and the data supports, that AI assistance erodes persistence via two mechanisms.

The first is hedonic adaptation. Once AI completes tasks in seconds, the reference point for how long a task should take shifts. Unaided work starts to feel counterfactually more effortful. Each act of offloading makes the next one more attractive. The process is self-reinforcing.

The second is loss of metacognitive calibration. Without opportunities to struggle independently, people never learn what they are capable of. They lose the self-knowledge that tells them they can finish a hard problem if they keep going. That self-knowledge is what sustains persistence under uncertainty.

These are not speculative mechanisms. They are well-documented features of how cognition interacts with tools. The paper’s contribution is showing that a ten-minute intervention is enough to start them turning.

---

The boiling frog, with a timer

The paper closes with this framing:

“This is analogous to the ‘boiling frog’ effect, where each incremental act feels costless, until the cumulative effect becomes overwhelming to address.”

The boiling frog framing has been used about AI risk before. The researchers in this paper have given it a timer. If ten minutes produces measurable persistence loss, sustained daily use over months will produce something much larger. By the time the loss is visible, it will be difficult to reverse.

This is what Slow AI has been saying since we started in July 2025.

Slow down.

Notice what is being traded. Use AI where it helps you think, not where it replaces the thinking. This research has now given the thesis empirical teeth.

It is also what the paper itself concludes:

“We hope that our work inspires the field to think about optimizing not just what people can do with AI, but what they can do without it.”

That is the question.

---

The Slow AI curriculum is built for the 27%. The ones who use AI for hints, not answers. Monthly live seminars, a community of over 250 educators, policymakers, and creators, and a CPD-accredited certificate in critical AI literacy. £100 a year.

---

A three-part framework for using AI so the scaffolding survives

This research names two specific mechanisms for persistence loss: hedonic adaptation (your reference point for task effort shifts) and loss of metacognitive calibration (you forget what you can do alone). A third mechanism, implicit in the data, is the direct-answer trap: 61% used AI to skip the thinking, and they were the ones it broke.

Each mechanism has a counter-practice. Here they are.

1. The Ten-Minute Floor

Counter-practice for: hedonic adaptation

Do not touch AI for the first ten minutes of any reasoning task. Set a timer if you need to. The goal is not productivity. The goal is to preserve your own sense of what the task feels like. Ten minutes of unaided struggle keeps your reference point anchored to your own effort rather than to the AI’s response time.

If you have been using AI for months without this floor, your reference point has already drifted. The first few times you do this will feel inefficient. That feeling is the recalibration.

The ten-minute floor is not anti-AI. It is anti-dependency. After ten minutes, use AI however you want. But the first ten belong to you.

2. The Hints-Not-Answers Rule

Counter-practice for: the direct-answer trap

Change your default prompt. Stop asking “what is the answer?” Start asking “what should I be thinking about?”

This is the single most important change the study’s findings suggest. The 27% who used AI for hints were fine. The 61% who used it for answers were not. The mechanism is not mysterious. When AI gives you the answer, you do not generate it. You receive it. When AI gives you a hint, you still have to do the work.

Examples of the switch in practice:

• Instead of “write me a paragraph about X,” try “what are the two or three strongest arguments a good paragraph on X would address?”

• Instead of “solve this problem,” try “what kind of approach would work here, and where do people usually get stuck?”

• Instead of “summarise this paper,” try “what is the one claim in this paper I should test against what I already know?”

The first prompt type generates an output. The second type generates scaffolding. The study says scaffolding is what protects persistence.

3. The Unaided Baseline Test

Counter-practice for: loss of metacognitive calibration

Once a month, do a task in your domain with no AI. A blog post, a problem set, a code review, a lesson plan. Whatever work you do, do some of it alone.

The point is not to prove you can. It is to remember what it feels like when you can. People lose persistence when they forget what struggle feels like on the inside. The unaided baseline test is how you stay calibrated.

Start with one hour a month. Scale up if it feels necessary. If you find the unaided work surprisingly hard, that is useful information. It means your baseline has slipped further than you thought and the test is doing its job

---

A diagnostic exercise to run this week

Pick a piece of work you would normally ask an AI toolto help with. Before you do, answer three questions:

1. Am I using AI to save effort or to learn something? If it is genuinely a save-effort task (expense report, summary of a meeting you attended), proceed without guilt. If it is a learn-something task (writing, reasoning, analysis), apply the ten-minute floor first.

2. Am I generating this argument or watching it generated? Test: could you write the next paragraph without the AI? If no, you are not scaffolding, you are watching. Stop and rewrite your last prompt as a hints prompt.

3. Can I reproduce this tomorrow without AI? If the answer is no, you did not learn it. You bought it. That is fine for some work and not for other work. Be honest about which is which.

---

Why this matters for you

Almost every reader of this post will use AI this week. The question is not whether to use it. The question is whether you will still be able to do the work without it in six months, in a year, in five years. The study suggests that answer is being decided in ten-minute increments you are not noticing.

The framework above is not complicated. Ten minutes. Hints, not answers. One hour of unaided work a month.

What the study makes newly clear is the cost of not bothering. Skip the framework and you are in the 61%. The 27% is a deliberate choice made every time you sit down at the keyboard.

Slow AI exists to help you make that choice.

Go slow.