A new Berkeley AI study found that about ten minutes of ChatGPT help can leave people quicker to quit hard problems once the chatbot is gone. Volunteers who had AI help started out more accurate, then skipped and missed more questions after it was taken away. Here’s what the researchers did, what it does and doesn’t prove, and how to use AI without losing your edge.
The findings were reported by Berkeley News on October 9, 2026.
Here are the quick facts, so you can skim first:
- Study title: “AI Assistance Reduces Persistence and Hurts Independent Performance.”
- Who ran it: Brian Christian of UC Berkeley’s Center for Human-Compatible AI, with scholars from Carnegie Mellon, MIT, Oxford and UCLA.
- Size: 1,222 people recruited online, across three randomized controlled trials.
- Tool tested: ChatGPT, with the option to ask it for answers.
- Status: peer-reviewed and presented at the Conference on Language Modeling (COLM).
Those five facts cover the core. The details below explain why the result matters for everyday use.
What the Berkeley AI study actually tested
The setup was simple. People worked through basic problems, and one group had ChatGPT open while the other worked alone. Partway through, the researchers removed the chatbot and watched what happened.
The team ran three versions of this. Two used fraction problems and one used SAT-style reading questions. The goal was to see whether leaning on AI changes how long you keep trying once it’s no longer there.
The paper first appeared as a preliminary draft in spring 2026. The version covered this week is the peer-reviewed update, which is available as an arXiv preprint.
What happened when the AI went away
The pattern was the same in all three experiments. The AI group began with better accuracy, then dropped off quickly after the tool was removed. They also skipped far more questions than people who had worked unaided.
This table sums up the three trials, based on the EurekAlert release:
| Experiment | Task | People | What happened |
|---|---|---|---|
| 1 | 15 basic fraction problems; AI removed after 12 | 354 | AI group started more accurate, then accuracy fell almost at once |
| 2 | Larger fraction replication | 667 | AI group gave wrong answers or quit; unaided group kept going |
| 3 | SAT-style reading comprehension | 201 | Persistence and accuracy both fell when the tool was removed |
The release doesn’t give exact percentages or effect sizes, so we can’t tell you how big the drop was. What it does say is that the direction held across all three tests.
Why quick answers can make hard work feel pointless
The researchers offer two explanations. First, instant AI answers make slow, careful work feel inefficient by comparison. Once you’ve seen how fast a chatbot can finish a problem, grinding through it yourself feels like a waste.
Second, AI removes what the team calls “productive struggle.” That’s the uncomfortable middle stretch where you’re stuck, and it’s also how people discover what they’re good at and what they enjoy.
Christian told Berkeley News he’s especially worried about science. He argues that over-reliance on AI could wear down habits like staying close to the data, immersing yourself in the material and working as a team.
How to use ChatGPT without losing your edge
The study doesn’t tell you to stop using AI. It does suggest a few habits worth trying. These are our suggestions, not the researchers’ rules:
- Try first. Spend a few minutes on the problem before you open the chatbot, even if your attempt is rough.
- Ask for a hint or a next step instead of the full answer.
- Ask ChatGPT to quiz you or explain its reasoning, so you do some of the thinking.
- Schedule short stretches of AI-free work on skills you want to keep sharp.
Christian’s own proposal points the same way. He suggests AI should default to acting like a tutor rather than an answer machine. He also keeps a daily notebook as a personal check against losing his own scholarly skills.
If you let kids or teens use chatbots for homework, our guide to the ChatGPT for Teens safety rating covers what parents should watch for.
What this study can’t tell you yet
Be careful with the headline. Participants were recruited online and worked on short, basic tasks. Neither the Berkeley release nor the EurekAlert summary says how long the effect lasts after the AI is removed.
That means this isn’t proof that ChatGPT permanently harms your thinking. It’s evidence that a brief stretch of easy help can change how hard you push right afterward. The ten-minute figure comes from the study’s coverage, and the model version used wasn’t specified.
The authors also describe an AI that coaches instead of answers as a better default, which is a design choice rather than a fixed outcome. If you want to see how fast chatbots are taking on more of your work, see our look at ChatGPT Finances going free in the US.
Frequently Asked Questions
Does using ChatGPT make you worse at thinking?
Not in any proven, lasting way. The study shows people gave up sooner and answered less accurately right after AI help was removed, but it didn’t measure long-term effects.
How many people were in the Berkeley AI study?
There were 1,222 participants, recruited online and split across three randomized controlled trials of 354, 667 and 201 people.
Was the study peer reviewed?
Yes. The updated paper was presented at the Conference on Language Modeling (COLM). An earlier draft circulated as a preprint in spring 2026.
Should I stop using AI for homework or work?
No, but try the problem yourself first and ask for hints rather than finished answers. That keeps some of the productive struggle the researchers say is valuable.
Our take: use AI as a coach, not a crutch
This is one study with short tasks, so don’t panic. But the pattern is easy to recognize from daily life: the easier the answer is to get, the less you want to work for it. Keep AI for the parts of a job you’re happy to hand off, and protect the skills you want to keep by doing the first attempt yourself.
Leave a Reply