Give people a chatbot for about ten minutes, take it away, and they quit sooner on the next hard problem. That is the finding of a paper on how AI reduces persistence by Grace Liu, Brian Christian and colleagues, built on three randomized trials with 1,222 participants.
The tasks were fractions and an SAT reading passage, a long way from quoting a renovation or reconciling a supplier statement. We think the result still belongs in every onboarding plan, because sooner or later every team works a day without the tool.
Key points
- What did the study actually test? Online participants solved problems either on their own or with ChatGPT available.
- Why does it matter to a business that AI reduces persistence? Because the tool goes away more often than plans assume.
- What should a team lead change in onboarding? Keep some unassisted work in the first weeks of every role.
- Is this really about jobs, or just fractions? Partly fractions. Participants were recruited online, the tasks were academic and the exposure was short.
What did the study actually test?
Online participants solved problems either on their own or with ChatGPT available. Partway through, the AI group lost access. After that, they solved fewer problems and skipped more, while the group that never had help stayed on task. The effect appeared after roughly ten minutes of AI use, and the paper was presented at a peer-reviewed conference this month.
In the first two experiments, people worked through 15 fraction problems and the AI group lost access after problem 12. The third used a reading comprehension task. UC Berkeley’s summary says the work was presented at the Conference on Language Modeling, and quotes Christian saying the result is about more than fractions and SAT questions.
Two limits matter. Neither the abstract nor Berkeley’s write-up gives effect sizes, so we can say the drop was there but not how large. And the authors offer their explanation as a hypothesis: that AI conditions people to expect immediate answers.
Why does it matter to a business that AI reduces persistence?
Because the tool goes away more often than plans assume. Outages, cut limits, withdrawn models and clients who ban AI on their file all leave staff working unaided. This study suggests the people most used to AI help are the first to stall on those days, so the plan has to cover them.
Because the tool goes away more often than plans assume. Vendors have outages, plans cut limits, a model gets withdrawn, a client bans AI on their file, or a task simply falls outside what the model handles. On each of those days your staff work unaided, and this study suggests the people most used to help are the first to stall.
The tool goes away in ordinary ways. A status page stayed green while apps failed. A vendor halved what a $200 plan included. A model was pulled from customers for nearly three weeks. Each of those was treated as a continuity problem. The Berkeley result adds a people problem underneath it.
New hires carry the most risk. Someone who learned the job with an assistant always on has never had to push through the hard part alone, which is why our piece on entry-level hiring matters here too.
What should a team lead change in onboarding?
Keep some unassisted work in the first weeks of every role. Have new staff do a short list of real tasks without AI, then the same tasks with it, and compare the results. The aim is to know who can finish the work when the tool is gone, and to build that ability before an outage tests it for you.
Checklist
- Pick three tasks per role that a person must be able to finish alone, such as a quote, a reconciliation or a reply to an angry customer.
- In the first two weeks, have new staff do them unaided first, then check their work with AI, and note how long each attempt took.
- Set the assistant to hint mode while people train. Ask it for the next step, not the answer, which is close to the tutor-style design Christian suggests.
- Run one hour a quarter with the tools off, on a normal workload, and note where people stall.
You can keep every tool you already pay for. The point is knowing what your team can still do when those tools are unavailable.
Is this really about jobs, or just fractions?
Partly fractions. Participants were recruited online, the tasks were academic and the exposure was short. Experienced workers may react differently, and months of use could build habits in either direction. We would act on it anyway, because the safeguard costs an hour and the first outage will show if you needed it.
Here is what would change our view. If a workplace study with real tasks and longer exposure finds no drop in persistence, unassisted practice goes from a safeguard to ordinary good training. Until that study exists, this is controlled evidence, and it points one way.
It also connects to a cost we raised in our column on AI slop grenades, where the hidden expense of AI output was the review time it pushed onto someone else. Here the hidden expense is ability, and it does not show up on any invoice. Training is the fix most firms skip, and the Canadian AI skills gap sits at 78%. Keeping skilled people matters too, which is the aim of the Mitacs plan on AI brain drain. Those lost days add up to the hidden time tax of AI productivity, which most teams never count.
Put one unassisted task into next week’s onboarding plan.
Frequently asked questions
Does using AI make people less persistent?
In three randomized trials with 1,222 online participants, people who used ChatGPT for about ten minutes solved fewer problems and skipped more once it was removed. The tasks were maths and reading, not workplace jobs.
Should businesses limit AI for new employees?
Limit it in training, without banning it. Keep a few core tasks that new staff learn to finish unaided first, and use hint-style prompts while they learn, so the team can still work during an outage.
Where was the AI persistence study published?
It is on arXiv as AI Assistance Reduces Persistence and Hurts Independent Performance, by Grace Liu, Brian Christian and colleagues, and was presented at the Conference on Language Modeling in October 2026, according to UC Berkeley.
Last fact-checked October 12, 2026. Sources are linked on the claims they support.
