The short version: A new study finds AI assistance reduces persistence: after about ten minutes of solving problems with ChatGPT, people who lost the tool solved fewer problems on their own and gave up more often than people who never had it. The headline stops there. The paper does not. Inside its second experiment, the people who used ChatGPT mainly for hints solved 76 percent of the unaided problems, against 77 percent for people who never had AI. The people who mainly took its answers solved 65 percent. How your crew asks matters more than whether they ask.
That split is the most useful line in the research for anyone training a new hire, and it is not in the press release. It sits in a usage analysis halfway through the paper, which we read on October 11, 2026. Below is what the study measured, where that split holds and where it is thin, and three habits that follow for an owner whose apprentice already has ChatGPT on their phone.
What did the AI persistence study actually test?
The paper is “AI Assistance Reduces Persistence and Hurts Independent Performance”, by Grace Liu of Carnegie Mellon, Brian Christian and Tsvetomira Dumbalska of Oxford, Michiel Bakker of MIT and Rachit Dubey of UCLA. It first appeared on arXiv in April, and Berkeley News reported on October 9 that the peer-reviewed version was being presented this week at the Conference on Language Modeling.
The design is simple, which is what makes it hard to wave away. The team recruited 1,222 adults through the research platform Prolific and ran three randomized experiments. One group worked a set of problems with GPT-5 open in a sidebar. The other group worked the same problems without it. Then the AI was taken away and both groups faced a short test.
- Fractions, first run: after 12 problems with AI, the AI group solved 57 percent of the test problems alone; the group that never had AI solved 73 percent. The AI group skipped 20 percent of test problems, the others 11 percent.
- Reading comprehension: on SAT-style passages, 76 percent solved against 89 percent, and 8 percent skipped against 1 percent.
- Fractions, larger replication: 585 people analyzed. The gap narrowed to 71 percent against 77 percent, and the difference in skipping was no longer statistically significant.
So the core finding replicates on accuracy, across math and reading, after roughly ten minutes of help. The authors’ explanation is that AI trains people to expect an immediate answer, and the slow work of figuring it out starts to feel like a waste of time.
Does it matter how people used ChatGPT?
In the larger replication, the researchers asked the AI group how they had used it. Sixty-one percent said mainly for direct answers. Twenty-seven percent said mainly for hints or clarifications. Twelve percent said they did not use it at all. Then the team compared each group’s test results.
| Group, larger fractions experiment | Share of AI group | Solved alone | Skipped |
|---|---|---|---|
| Used AI mainly for answers | 61% | 65% | 13% |
| Used AI mainly for hints | 27% | 76% | 5% |
| Had AI, did not use it | 12% | 89% | not reported |
| Never had AI (control) | n/a | 77% | 7% |
Read the first two rows together. Same experiment, same problems, same tool, and nobody’s pay depended on getting answers right. The people who asked ChatGPT for hints kept going when it disappeared; the people who took its answers skipped problems more than twice as often and solved 11 points fewer. The hint group’s numbers sit almost exactly on the control group’s.
Now the honest limits, because this is where a summary would overreach. The usage groups were self-reported and not randomized, and the authors say plainly that this part is not causal evidence. The paper compares the hint group with the answer group, not with the control group, so “as good as never having AI” is our reading of the numbers, not a tested result. And a person who chooses to ask for hints may already be the kind of person who sticks with a hard problem. The authors checked one thing that helps: the three usage groups did not differ on a pretest taken before anyone touched the AI. That rules out the simplest version of “the hint people were just better at fractions.”
The authors also think the fix belongs to the companies that build these models, and they call user-side workarounds such as a tutoring mode “band-aids.” That is a fair point about the industry. For a business owner on a Tuesday, a band-aid you can apply this week beats a redesign you are waiting on. Christian put the larger goal in terms we would sign: human abilities should be “augmented rather than supplanted.”
What should an owner do when a new hire uses ChatGPT?
Nothing in this study says to take the tool away. The people who used it well did fine, and AI that answers questions at 6 a.m. on a job site is real capacity for a small crew. It says to be deliberate about which kind of help goes with which kind of work. Three habits follow.
1. Split the work into “learning” and “doing.” The study measured learning: whether a skill stuck once the help was gone. When your estimator needs a quote out the door, an answer is the right kind of help. When your second-year apprentice is learning to size a drain line or read a panel schedule, it is not. Name the three or four skills each new hire must eventually do without a phone, and treat those differently. Our earlier piece on how an AI tutor matched an expert human tutor found the same pattern from the other side: the tutoring that worked started from the learner’s own mistakes, not from the answer.
2. Turn on the hint setting the tools already have. ChatGPT’s study mode, per OpenAI’s help page read on October 11, 2026, is available on every ChatGPT plan on web, iPhone and Android: type @study in the message box or choose Study from the + menu. One catch worth knowing before you build a crew setup around it: the page says study mode is not available in Temporary Chats, custom GPTs or Projects, so if your team works inside a shared Project, it has to start a regular chat for the learning work. OpenAI also notes it may still hand over a direct answer at times. Google’s Gemini app has a similar Guided Learning mode, launched in August 2025. In Claude, or any assistant, a plain first line does the same job: “Don’t give me the answer. Give me the next step, then wait.”
3. Run a one-problem check, no phone. The study’s test was three problems without the AI. Copy it at a smaller scale: once a week, hand the new hire one real problem from their learning list and ten minutes without the assistant. No grade, no lecture, just a look at whether the skill is landing. If it is not, the fix is usually the habit in step 2, not a ban. It fits neatly into a 30-day learning plan or the standard work instructions you already hand new people.
One more thing the numbers say quietly. The 12 percent who had ChatGPT open and chose not to use it solved 89 percent of the test problems, the best result of any group. We would not build a policy on a self-selected group of 37 people. But it is a decent reminder that having the tool and leaning on the tool are two different things, and the choice between them belongs to the person, not the software.
Frequently Asked Questions
Does using AI make you worse at your job?
Not in general, according to this study. It measured something narrower: after about ten minutes of solving problems with ChatGPT, people who then lost the tool solved fewer problems alone and skipped more. That is a learning effect. Using AI to get finished work out the door is a different question the study did not test.
Is the AI persistence study peer-reviewed?
Berkeley News reported on October 9, 2026 that the peer-reviewed version was presented at the Conference on Language Modeling. The paper, by researchers from Carnegie Mellon, Oxford, MIT and UCLA, has been public on arXiv since April 2026 under the number 2604.04721.
Did people who asked ChatGPT for hints do better?
In the larger fractions experiment, people who said they mainly used ChatGPT for hints solved 76 percent of the test problems alone and skipped 5 percent, against 65 percent and 13 percent for those who mainly took answers. The groups were self-reported, so the authors do not treat this as causal.
How do I make ChatGPT give hints instead of answers?
Use study mode: type @study in the message box or pick Study from the + menu. OpenAI’s help page says it works on every plan but not in Temporary Chats, custom GPTs or Projects. In any assistant, asking for the next step rather than the answer does much the same.
When someone on your crew is learning a new skill, do they ask the AI for the answer or for the next step, and have you ever asked them which?
