Ten Minutes With an AI Assistant Is Enough to Erode People’s Persistence at Hard Problems, Major Study Finds

Just ten minutes of relying on an AI assistant is enough to measurably erode a person’s ability to focus on and persist through difficult problems, according to a peer-reviewed study presented this week at the Conference on Language Modeling. The research, a collaboration between scholars at Carnegie Mellon University, the University of Oxford, the Massachusetts Institute of Technology and the University of California, Los Angeles, offers some of the clearest causal evidence yet that the way today’s AI systems are designed to “help” may be quietly undermining the very skills people turn to them to build.

The paper, titled “AI Assistance Reduces Persistence and Hurts Independent Performance,” was led by Brian Christian, a research fellow with UC Berkeley’s Center for Human-Compatible AI and the author of the widely praised 2020 book The Alignment Problem. An early draft of the work circulated earlier this year and attracted attention from outlets including WIRED and Psychology Today; the updated, peer-reviewed version was presented at the Conference on Language Modeling during the week of October 5. Last week, The New York Times published a story about the paper and the broader body of similar research, according to Berkeley News.

To test how quickly human persistence could deteriorate under AI assistance, the team designed a series of randomised controlled trials and recruited 1,222 participants online. In the first experiment, involving 354 participants, one group was asked to solve 15 basic fraction problems entirely on their own, while a second group had ChatGPT open on the side — free to consult it for help or even to ask for the answer outright. The AI-assisted group started off more accurate than the unassisted group. Then, after 12 problems, the researchers took the tool away. Almost immediately, the people in the formerly assisted group stopped solving questions accurately, while the control group’s performance held steady and even improved.

The team then repeated the experiment with a second, larger group of 667 participants. The result was the same, and starker: where the AI users got answers wrong or gave up entirely once the assistance was withdrawn, the group that had worked without AI from the start stuck with the task and finished more successfully. A third experiment, with 201 participants, tested whether the effect extended beyond arithmetic. Using SAT-style reading comprehension questions, the researchers again found that once the AI tool was removed, both persistence and accuracy dropped.

“You almost couldn’t ask for a clearer story,” Christian told Berkeley News.

The charts tell the tale visually: the solve rate for the AI-assisted group, plotted as a line, drops sharply downward right after the 12th problem — the point at which the tool disappears — while the line for the group that never had AI holds steady. A second graph, tracking how often participants simply skipped questions, shows the inverse: the skip rate for the AI group shoots up after the tool is removed, a behavioural signal that the participants were not merely making mistakes but giving up.

The authors argue that the mechanism is motivational as much as cognitive. Current AI systems are, in the paper’s framing, fundamentally short-sighted collaborators: they are optimised to deliver instant, complete answers and almost never say no, the way a human mentor or tutor would. When answers arrive in seconds, the brain recalibrates what a reasonable amount of effort feels like. Anything slower starts to feel inefficient, even punishing. Persistence — the willingness to sit with a hard problem — is, the researchers note, foundational to skill acquisition and one of the strongest predictors of long-term learning. And it is precisely persistence that the assistance erodes.

The study’s team calls what is lost “productive struggle”: the uncomfortable but formative experience of working through a challenge on one’s own, which is how people discover their strengths and, as the researchers put it, find their passions. When the machine absorbs the struggle, the learning goes with it.

“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,” Christian said. “I experience it too.” The researcher, who has spent two decades studying how artificial intelligence will change minds, described himself as sometimes euphoric about the technology’s advances — and simultaneously worried that too much time working alongside AI could cost him his own scholarly edge. He still writes in a paper notebook every day, he said, a practice that keeps him grounded.

The implications extend well beyond fractions and reading tests. Christian argued that the dynamic is especially dangerous in research and academia, where the value of a scientist lies not in producing answers but in staying close to data, being immersed in the material, and working as part of a team. He warned that something significant could be lost when a team of ten people developing the ideas behind a breakthrough is whittled down to two or three researchers and a chatbot. “This is not a story about fractions and SAT problems,” he said. “This is something that is happening to human knowledge and human expertise from top to bottom and from grade-school students all the way to the leading experts in the world.”

The finding lands in the middle of an already heated debate over AI in classrooms. In early September, New York City announced a one-year moratorium on student-facing generative AI from pre-kindergarten through eighth grade — affecting roughly 600,000 students, about two-thirds of the city’s public school enrolment — while Los Angeles blocked AI on school devices indefinitely for more than 378,000 students. The cities framed the moves as an effort to protect children’s learning from precisely the kind of dependency the Berkeley study now documents with experimental rigour. Critics of the bans argue that keeping students away from AI entirely merely hands the advantage to children with tech-savvy parents, and that the students who most need AI literacy will be the last to get it.

Even the companies building the technology have begun to concede the point. Frontier AI firms have acknowledged that offloading tasks to chatbots can reduce a person’s baseline understanding and mastery of a subject. Christian’s prescription is architectural rather than moral: instead of defaulting to quick answers, he argued, AI responses could default to being more instructive — behaving less like an oracle and more like a tutor that scaffolds understanding. “We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted,” he said.

“We’ve built these systems to be helpful, but are they being helpful to us in a way that’s actually helping us?” Christian asked. “We showed, ironically, that they are often helping us in ways that are kind of unhelpful.”

Analysis: Why It Matters

The Berkeley paper matters less for what it discovered than for what it proved. Anecdotes about AI making students lazy, or surveys showing professionals leaning on chatbots, are easy to dismiss as nostalgia or moral panic. A randomised controlled trial with 1,222 participants, replicated across three experiments and two domains, is much harder to wave away. This is causal evidence — not correlation — that brief exposure to AI assistance changes how people behave when the assistance is gone. The study moves the deskilling debate out of the realm of opinion and into the realm of measurement.

The ten-minute figure is the detail that should give product designers pause. Earlier assumptions placed the danger of AI dependency in the slow lane: a gradual drift over months of use, an erosion so gentle it would be imperceptible until the skill was gone. The Berkeley results relocate the risk to the fast lane. Ten minutes — a single homework session, one debugging detour — was sufficient to alter both performance and motivation. If the effect is acute rather than cumulative, then every design choice that encourages instant answers is a design choice with immediate cognitive cost. The skip-rate finding reinforces this: the AI group did not just perform worse, they gave up more often. Quitting is not a knowledge gap; it is a motivation gap. The tool did not merely answer their questions — it reshaped their relationship with effort itself.

That reframing has direct consequences for the two most live policy debates in AI. The first is education. New York City’s moratorium and Los Angeles’s indefinite ban were enacted on intuition: that children learn by struggling and that chatbots short-circuit that process. The study supplies the empirical ammunition those policies lacked — but it also exposes their bluntness. A ban treats all AI use as answer-getting, while the research points toward a more precise target: the mode of assistance. Christian’s tutor-mode proposal — AI that instructs rather than answers — suggests regulation and design could aim at how the tool behaves rather than whether it exists. The question for school districts is no longer simply “ban or embrace,” but “under what interaction design does the tool preserve productive struggle?”

The second debate is in the enterprise, where the stakes are commercial rather than pedagogical but the mechanism is identical. AI agents are being sold as autonomous colleagues that plan work, write code, and operate across applications. Every task an agent completes invisibly is a task the human never practised. The study implies that organisations deploying these systems should be measuring not just output but retained human capability — whether the people in the loop can still do the work when the system goes down. The September multi-platform outage, when ChatGPT, Claude and Grok all stumbled within hours of each other, offered a preview of what dependency without capability looks like. The Berkeley finding suggests the capability erosion can begin within a single afternoon.

There is also a subtler implication for the AI labs themselves. The paper’s framing — AI systems as “short-sighted collaborators” optimised for instant answers — is a direct challenge to the helpfulness metric that governs model development. Helpfulness, as currently optimised, means minimising time-to-answer. But the study suggests that the fastest answer is not the most helpful one, at least not for the human on the other end. If labs took the persistence effect seriously, the next frontier of model evaluation would not be benchmarks of accuracy but benchmarks of human growth: does working with this system leave the user more capable than before? That is a harder thing to measure than a test score — but it is, the researchers would argue, the thing that actually matters.

What to watch next

• Whether other labs replicate the finding. A single paper, however well designed, invites confirmation — watch for replications with different tasks, age groups and AI systems, and for studies testing whether the effect persists over days rather than minutes.
• Whether “tutor mode” survives contact with reality. Christian’s prescription — AI that instructs instead of answering — sounds straightforward, but it runs against the commercial incentives of engagement: users like fast answers. Watch whether any major lab ships a persistence-preserving mode and, critically, whether users choose it.
• The school-policy experiments. New York’s moratorium runs through the school year, with a coalition expected to issue recommendations in April 2027. Watch for early data on whether the ban changes learning outcomes — and whether the Berkeley study becomes the citation that justifies or refines it.
• Enterprise capability audits. As AI agents take over more workflows, watch for the first large employers to formally measure retained human skill — or the first incident in which an outage exposes a workforce that can no longer do its own work.

Sources

• UC Berkeley News — “Using AI for just 10 minutes erodes your ability to persist at hard things” (Oct 9, 2026)
• arXiv — “AI Assistance Reduces Persistence and Hurts Independent Performance” (the full paper)
• bioengineer.org — “Ten Minutes With AI Can Weaken Your Ability to Persist Through Hard Problems”
• AI Weekly — “Berkeley study: 10 minutes of AI use erodes persistence”