Ten minutes. That’s all it took. A randomized controlled study out of UC Berkeley found that a brief stretch of ChatGPT assistance was enough to measurably damage participants’ persistence — how long people stuck with a hard problem once the AI got taken away. Not a survey. Not a vibe. A controlled experiment, published October 9, 2026, with the kind of causal design that’s been mostly absent from the “AI is making us dumber” conversation up to now.

We’ve all heard the anecdotal version of this story — someone opens with a confession that they can’t write an email anymore without a draft from the model, or a teacher complaining that students give up the second autocomplete stops helping. This is the first time anyone’s run it as an actual experiment, with a control group, randomization, and a removal condition designed specifically to catch the effect in the act.

Quick Summary: What Happened

Detail	Info
Published	October 9, 2026, on arXiv, covered by Berkeley News and EurekAlert
Who ran it	Brian Christian at UC Berkeley’s Center for Human-Compatible AI, with co-authors from Carnegie Mellon, MIT, Oxford, and UCLA
Study size	1,222 participants across three randomized experiments
Tasks	Fraction problems (two experiments, 354 and 667 participants) and SAT-style reading comprehension (201 participants)
The twist	ChatGPT access was pulled from the AI-assisted group partway through (around problem 12); the control group never had it
The result	AI-assisted accuracy and persistence fell off sharply once the tool was gone. The no-AI control group held steady — and in some runs, kept improving
Presented at	The Conference on Language Modeling (COLM 2026), San Francisco

Bottom line: This isn’t a hunch about “AI dependency” anymore. It’s a controlled trial showing that a few minutes of leaning on ChatGPT changes how you perform the moment it’s gone — and the people who never got the help did better once things got hard.

How ChatGPT Use Erodes Persistence

Here’s the design, because it matters more than the headline. Researchers split participants into an AI-assisted group and a no-AI control group, then had both work through a set of problems — basic fraction arithmetic in two of the three experiments, SAT-style reading comprehension in the third. The AI group got ChatGPT access for the early problems. Then, partway through — around problem 12 — the researchers cut it off. No warning baked into the task, just: the tool you’d been leaning on is gone now, finish the rest yourself.

The control group never had AI access at all. Same problems, same time, same difficulty curve. The only variable that moved was whether someone got a few minutes of assistance before being left to it.

What happened next is the whole study. The AI-assisted group’s solve rate and accuracy dropped hard once the assistance disappeared — accuracy fell by double digits in percentage-point terms in more than one of the three experiments. Skip rates went up too: people in the AI-assisted group gave up on problems at meaningfully higher rates than the control group once they were on their own. Meanwhile the group that never had AI in the first place didn’t just hold steady. In some of the runs, they kept getting better as the task went on — exactly the kind of improvement-through-struggle pattern you’d expect from people building a skill under pressure, rather than borrowing an answer.

That’s the part worth sitting with. This isn’t “people who use AI are worse at fractions.” It’s “the exact same person, randomly assigned, does worse on the same problems after ten minutes of AI help than if they’d never gotten the help at all.” The AI didn’t just fail to teach them. By the study’s own framing, it actively eroded a capacity they’d have kept if they’d been left alone.

Why Researchers Think This Happens

1. Immediate answers remove the discomfort that drives learning. Struggling with a hard problem is unpleasant, and that discomfort is doing real cognitive work — it’s what pushes people toward strategies that stick. Remove the discomfort and you remove the mechanism.
2. People stop building their own problem-solving routine. When the AI is handling the “how,” users never develop their own path through a problem type, so there’s nothing to fall back on when the AI disappears.
3. Short exposure is enough to shift expectations. The effects showed up after roughly ten minutes of use — not months of chronic dependence. The threshold for behavior change is lower than most of us assumed.
4. Persistence itself is the thing that erodes, not just knowledge. The control group wasn’t smarter about fractions. It had more practice at pushing through a hard problem instead of quitting on it — and that’s precisely the muscle the AI-assisted group lost the chance to build.

Brian Christian, the study’s lead researcher at Berkeley’s Center for Human-Compatible AI, put it in terms that land harder than a stats table: “It’s a striking finding, but in a way it supplies ammunition for a story that a lot of us kind of feel in our gut.” He followed that with the question the paper is really asking: “We’ve built these systems to be helpful, but are they being helpful to us in a way that’s actually helping us?”

Why This Matters

This site covers tools for people who use AI as part of their actual job, daily, not as a curiosity. That’s exactly the population this study should worry. If ten minutes of assisted problem-solving is enough to measurably dent independent performance in a lab setting, the realistic exposure most knowledge workers get — hours a day, every day, across every kind of cognitive task — isn’t a scaled-up version of the same effect. It’s a different animal we don’t have data on yet, because nobody’s run the year-long version of this experiment.

We’ve written before about getting started with AI tools and the best AI coding assistants with an implicit assumption running underneath most of that coverage: that leaning on AI for the easy parts frees you up for the hard parts. This study doesn’t disprove that for every task. But it’s the first causal evidence that the opposite can also be true — that leaning on AI for the hard parts erodes your ability to do them at all, not just your willingness.

It’s also a pointed footnote to the ChatGPT 5 review we ran after months of daily use: a tool can be genuinely excellent at producing a correct answer and still be a net negative for the skill you were trying to build in the first place. Those aren’t contradictory claims. They’re the same tool evaluated on two different axes — output quality versus what using it does to you — and this study is squarely about the second one.

The education angle is the most immediate. We’ve covered how AI is reshaping classrooms in terms of personalization and faster feedback loops. This study is a direct counterweight to that optimism: if a ten-minute dose of assistance measurably reduces a college-age adult’s persistence on a problem set, the stakes for a student still building foundational math or reading skills are higher, not lower.

What Are Your Options Now

If you manage people who use AI tools daily, build in unassisted stretches. The study’s design is almost a prescription on its own: performance held up fine for the group that never had the crutch. If part of someone’s job is a skill you want them to actually retain — debugging without a coding assistant, drafting without a writing model, estimating without a spreadsheet formula generator — some portion of their practice needs to happen with the AI switched off, not just at the end of a project when the stakes are highest.

If you’re evaluating AI tools for your team, separate “makes this task faster” from “builds this skill.” Our AI safety guide for business already makes the case for vendor diligence beyond the sales pitch. Add this to the list of questions worth asking before you roll out an assistant across a whole department: what happens to the team’s competence on this task if the tool goes down, loses budget, or gets pulled mid-project?

If you’re an individual user, treat AI removal as a planned event, not a crisis. The study’s cleanest signal is that the control group — the one that never had AI at all — did fine, and in places improved. That’s not an argument for abstinence. It’s an argument for occasionally doing the hard version of your own job on purpose, the way you’d deadlift without a spotter some days just to know you still can.

The Bigger Picture

This lands in the middle of a year where AI trend coverage has mostly been about capability — what models can now do that they couldn’t six months ago. This study is one of the first big, methodologically serious pieces of evidence pointed the other way: at what sustained reliance does to the human using the tool, independent of how good the tool gets. Those are different research questions, and this site has spent more time on the first one than the second. That balance is probably overdue for a correction.

It also arrives as AI companies lean harder on language like “thought partner” and “copilot” — framing that implicitly promises the tool makes you better at the underlying skill, not just faster at the task. This study is the first randomized evidence that, at least for persistence on hard problems, the opposite can be true inside a ten-minute window. Christian’s own framing — a finding that “supplies ammunition for a story that a lot of us kind of feel in our gut” — is a fair description of why this is landing the way it is. People have suspected this for a while. Now there’s a controlled trial instead of a shared feeling.

One caveat worth being honest about: this is lab data on fraction problems and reading comprehension, not a longitudinal study of professionals using AI at work for years. The study measures what it measures — short-term persistence after brief, controlled exposure. Whether the same mechanism scales to a software engineer who’s used an AI coding assistant daily for three years is an inference, not a finding, and the researchers haven’t claimed otherwise.

Our Take

We think this is the most important piece of AI research we’ve covered this year that has nothing to do with a product launch. Capability benchmarks get refreshed every few weeks. Causal evidence about what AI assistance does to the human operating it is rare, and this is a clean, randomized version of it — not a survey asking people how dependent they feel, which is the kind of data this debate has run on until now.

What we’d flag above the headline number: the control group didn’t just avoid the drop. In multiple runs, it kept improving as the task got harder. That’s the detail that should bother daily AI users more than the accuracy drop itself. The question isn’t just “does AI help make me worse without it.” It’s “what am I not getting better at, right now, because the AI keeps answering before I have to try.”

Frequently Asked Questions

What exactly did the Berkeley study measure?

It measured persistence (whether people kept trying or gave up on a problem) and independent accuracy (how well people performed once AI help was removed), across three randomized experiments covering fraction arithmetic and SAT-style reading comprehension, with 1,222 total participants.

How much AI use caused the effect?

Roughly ten minutes — the researchers cut off ChatGPT access partway through the task, around problem 12, after a relatively brief assistance window. The effect wasn’t the product of months of chronic use; it showed up fast.

Did the control group do better than the AI-assisted group?

Yes. Once AI access was removed from the assisted group, their accuracy and persistence dropped sharply. The control group, which never had AI access, performed steadily throughout and in some experiments kept improving as the task progressed.

Who conducted the research?

Brian Christian, a researcher affiliated with UC Berkeley’s Center for Human-Compatible AI, led the study with co-authors from Carnegie Mellon, MIT, Oxford, and UCLA. It was presented at the Conference on Language Modeling (COLM 2026) in San Francisco.

Does this mean I should stop using ChatGPT for hard tasks?

Not necessarily. The study doesn’t argue AI assistance is useless — it argues that unassisted practice builds persistence AI assistance doesn’t. The practical takeaway is balance: use AI when speed matters, but keep some practice unassisted when the underlying skill matters too.

Is this the same thing as “cognitive offloading” research from earlier years?

It’s related but more rigorous. Earlier cognitive-offloading work on AI tools has largely been correlational — surveys asking people how much they rely on AI and how confident they feel. This study uses random assignment and a removal condition, which is what lets it claim the relationship is causal rather than just associated.

Where can I read the actual paper?

The full paper, “AI Assistance Reduces Persistence and Hurts Independent Performance,” is posted on arXiv.

---

Last updated: October 11, 2026. Sources: UC Berkeley News — Using AI for just 10 minutes erodes your ability to persist at hard things · arXiv — AI Assistance Reduces Persistence and Hurts Independent Performance · EurekAlert — Using AI for just 10 minutes erodes your ability to persist at hard things · Conference on Language Modeling (COLM 2026).

Related reading: ChatGPT 5 Review 2026 · AI Safety for Business: What Leaders Need to Know · Best AI Tools for Education in 2026 · Getting Started with AI Tools · Best AI Coding Assistants in 2026 · AI Trends in 2026: What’s Actually Changing