Estimated reading time: 5 minutes · Last updated:
A study co-authored by Brian Christian finds that just 10 minutes of active AI assistance can reduce a person’s ability to persist on difficult tasks. The team ran randomized controlled trials with 1,222 participants in total and reported the results at the Conference on Language Modeling; the experiments included groups of 354, 667 and 201 people. When researchers removed access to the AI tools, accuracy fell and skipping rose sharply among the participants who had relied on AI, suggesting a rapid erosion of what the authors and team described as the capacity for 'productive struggle'.
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,
Brian Christian, research fellow, UC Berkeley Center for Human-Compatible AI
Key takeaways
- Core finding: A study presented at the Conference on Language Modeling found that 10 minutes of AI assistance sharply reduced users' persistence on hard tasks.
- Sample size: Researchers ran randomized controlled trials with 1,222 participants in total, including experiments of 354, 667 and 201 people.
- Task details: The first experiment asked participants to solve 15 fraction problems and removed AI access after 12 problems; accuracy for prior AI users collapsed after removal.
- Authorship and collaborators: Co-author Brian Christian and teammates from UCLA, the University of Oxford, Carnegie Mellon University and the Massachusetts Institute of Technology reported the results, as first reported by University of California, Berkeley.
Table of contents
How the researchers tested brief AI use
The team designed randomized controlled trials to measure short-term effects of AI assistance on focus and task persistence. Across three experiments the total sample was 1,222 participants recruited online. The first trial had 354 participants working on 15 basic fraction problems; in that trial researchers allowed one group to use an AI assistant while another group worked without it. After 12 problems the AI was removed from the assisted group and performance was tracked for the remaining items. Two further experiments—one with 667 participants on similar arithmetic tasks and one with 201 participants using an SAT reading comprehension prompt—used the same remove-access design.
Presentation and documentation of the work appeared at the Conference on Language Modeling and in a draft posted to arXiv titled "AI Assistance Reduces Persistence and Hurts Independent Performance." The authors used the remove-access step to replicate the real-world moment when a person can no longer consult a tool and must continue on their own, making the setup a close test of short-term dependence rather than long-run learning.
What happened when AI help disappeared
When the researchers cut off access to the AI assistant, the assisted participants showed two consistent responses: accuracy dropped and the rate of skipping or giving up rose. In the first experiment the assisted group began more accurate than the control group while the AI was available, but after removal that group’s solve rate plunged. The second experiment returned a similar pattern, and the third—using an SAT-style reading item with 201 participants—echoed the earlier results. The rapid change occurred within the same session, indicating a near-immediate behavioural shift rather than a gradual one.
The authors frame the change as a loss of productive effort: when a tool provides answers instantly, tasks that require longer cognitive effort start to feel unusually costly, so people stop persisting. That mechanism helps explain why brief exposure—ten minutes of active use in the study—was sufficient to change behaviour on subsequent problems.
Why 'productive struggle' matters for skill formation
The team uses the term productive struggle to name a learning process in which difficulty and effort reveal misunderstandings and drive deeper engagement. In education research, productive struggle is associated with long-term retention and the discovery of problem-solving approaches; removing it can mean learners never develop the mental habits that sustain independent performance. The paper argues that instant solutions from an assistant can short-circuit those habits, replacing trial-and-error and reflection with fast answers.
Brian Christian framed the concern beyond classroom arithmetic: he warned that if researchers and professionals outsource cognitive work routinely, the social processes that produce expertise—immersion in data, iterative debate and shared problem-solving—could shrink. The authors note that frontier AI companies have acknowledged task offloading can reduce baseline mastery, and they suggest design and policy responses that would make AI more instructive when the goal is learning rather than mere speed.
Practical implications for educators and research teams
For instructors, the experiments imply that giving learners unsupervised, answer-producing AI access can harm persistence on subsequent independent work. The study's remove-access protocol models the common classroom and exam moment when learners must proceed without tools, so the findings argue for careful scaffolding: prompts and interfaces that push explanation and reflection rather than only answers. For research teams, the concern is that offloading idea-generation or routine reasoning to assistants may shrink the collaborative effort that often yields breakthroughs.
The authors, together with collaborators at the University of Oxford, UCLA, Carnegie Mellon University and the Massachusetts Institute of Technology, recommend alternatives such as tutor-like responses from assistants, interface designs that require users to explain or justify an answer, and evaluation regimes that measure independent problem-solving alongside assisted productivity. Those proposals aim to preserve the speed benefits of AI while protecting the cognitive practice that builds expertise.
| Experiment | Participants | Task | Design detail | Outcome |
|---|---|---|---|---|
| Experiment 1 | 354 | 15 fraction problems | AI available, removed after 12 problems | Accuracy collapsed and skipping rose after removal |
| Experiment 2 | 667 | Basic fraction problems | AI available then removed | Similar drop in persistence and accuracy |
| Experiment 3 | 201 | SAT reading comprehension prompt | AI available then removed | Reduced persistence and lower independent accuracy |
How this could play out
The case for
- Design changes that make assistants more instructional—requiring explanations or giving stepwise hints—could retain speed while protecting learning.
- Adopting evaluation that measures independent performance alongside assisted output would provide incentives to preserve core skills.
The case against
- Widespread default use of answer-producing assistants could shorten the time learners spend in productive struggle, reducing long-term mastery across education levels.
- Research teams that rely heavily on assistants risk losing deep familiarity with data and methods, which could lower the capacity for novel discoveries.
What to be careful about
- Reduced independent problem-solving skills in students when assessment still requires unaided performance.
- Erosion of research teams' collective expertise if routine reasoning is consistently delegated to AI.
- Uneven impacts across institutions: those with tighter instructional design may preserve skills while others do not, widening skill gaps.
The bottom line
The paper co-authored by Brian Christian provides a clear experimental signal that brief, answer-focused AI assistance can change how people approach subsequent, unaided work. The remove-access design—used across trials with 354, 667 and 201 participants and reported to a total sample of 1,222—shows an immediate behavioural shift: accuracy and persistence fall when tools are no longer available. That does not rule out longer-term benefits if assistants are deliberately designed to teach; it does, however, put the burden on designers and educators to choose interfaces and assessment strategies that preserve productive struggle and independent skill.
What to watch
- Watch for journal publication or formal proceedings of the Conference on Language Modeling summarizing the peer-reviewed paper; no date has been set.
- Watch for replication studies of the arXiv paper's remove-access experiments; no date has been set.
Frequently asked questions
How long did participants use AI before effects appeared?
The study reports that roughly 10 minutes of active AI assistance preceded the observed decline in persistence; this effect showed up within the same session when the tool was removed.
Which tasks did the researchers test?
The experiments included arithmetic tasks—one protocol used 15 fraction problems—and an SAT-style reading comprehension prompt; the trials sampled 354, 667 and 201 participants respectively.
Who ran the study and where was it presented?
Brian Christian is a co-author; the paper names collaborators from Carnegie Mellon University, the University of Oxford, UCLA and the Massachusetts Institute of Technology, and it was presented at the Conference on Language Modeling.
Related reading
