A few minutes of AI assistance can make a difficult task feel easier. The catch, according to research involving 1,222 participants, is what happens when that assistance disappears. In controlled maths and reading exercises, people who had been using an AI tool became more likely to stop trying and performed worse once it was removed. The finding does not show permanent damage to anyone’s ability to think, but it raises a practical question about using instant answers while learning.
The study, reported by UC Berkeley in October 2026, used randomised experiments rather than simply asking people whether they liked chatbots. Some participants tackled problems alone; others could consult AI and then had to continue without it. The pattern in the later questions mattered as much as the initial scores. Assistance made early work smoother, but some users struggled to persist when the shortcut was taken away.
How AI assistance changed the experiment
Imagine working through a set of fractions. For the first few questions, there is a helpful assistant beside you that can show a method or simply produce the answer. Then, without warning, it disappears and you must finish on your own. Your underlying knowledge has not necessarily vanished. But the task may suddenly seem more demanding than it did a minute earlier, and you may feel less inclined to persevere.
That is close to the situation the research team designed. In one trial, 354 participants worked on a sequence of fraction problems. The AI-assisted group could use ChatGPT during the earlier part of the activity, while another group worked unaided. Early answers from the AI-assisted group were more accurate. When access was removed towards the end, their performance deteriorated sharply relative to the group that had practised without the tool.
A second experiment involved 667 participants, and a third tested reading comprehension with 201. Across the programme, researchers examined not only whether answers were correct but whether people continued trying or skipped difficult items. The headline result was a drop in unassisted performance and persistence after AI was withdrawn. It is evidence about short experimental sessions, not a direct test of what happens after months of professional or educational use.
The full research paper, AI Assistance Reduces Persistence and Hurts Independent Performance, was revised in October 2026 and presented at the Conference on Language Modeling. Its authors frame the problem as a mismatch between immediate task success and developing the willingness to work through difficulty. A tool can be very good at delivering an answer without necessarily helping the person grow more capable of finding an answer independently.
The missing ingredient may be the struggle itself
Many forms of learning are uncomfortable. A new language takes repeated attempts, arithmetic requires practice and an unfamiliar piece of software can feel maddening before its controls become natural. The discomfort is not always wasted time. It can be the period in which somebody tests ideas, notices mistakes and builds a method that works the next time. When a tool removes the difficult step too quickly, it may also remove the chance to practise that persistence.
That does not make all struggle desirable. Nobody needs to waste an afternoon on a routine calculation that could be checked instantly. The distinction is between tasks where the result is the aim and tasks where the process is part of the aim. A professional preparing a report may reasonably use AI to reduce repetitive work. A student learning the reasoning behind an answer may need to attempt the problem first, even if that takes longer.
The study helps explain why a short burst of excellent assisted performance can be deceptive. A learner might finish more questions, experience fewer frustrations and feel more competent, yet find the next unaided problem harder. Teachers and managers should therefore be cautious about judging learning by the quality of AI-assisted output alone. A polished answer does not always reveal who understood the process or how readily they could repeat it without help.
A different LiveAIWire report has explored whether AI assistance can leave human skills weaker. That work asks related questions through a separate experimental lens. The Berkeley-linked research adds a particularly accessible observation: removing the assistant halfway through a task reveals a contrast that would be missed if researchers looked only at how well users performed while the tool remained available.
Why the study does not prove permanent decline
The striking results have limits that should stay attached to them. Participants were recruited online and asked to perform specific maths and reading tasks over relatively short periods. Their behaviour in those sessions cannot automatically be extended to every worker, student, artist or software developer. Nor did the experiments establish that the cognitive effect is permanent, that using AI occasionally causes a lasting reduction in ability, or that the participants could never recover with practice.
The study was designed to measure a short-term cause-and-effect relationship under controlled conditions. That strengthens its conclusions about the tasks tested, but the controlled setting is also what limits generalisation. Long-term studies could ask whether people develop better ways of working with assistance, whether different kinds of prompts change the outcome, and how much independent practice is needed to retain skills. Those questions remain open.
It is also possible to imagine an AI system behaving differently. An assistant might provide a hint rather than an answer, ask the learner to explain a step or deliberately withhold the solution until they try. That approach could preserve some of the productive difficulty the researchers believe matters. The experiments should not be read as proving that every design of AI tutor must have the same effect as the assistance used in these tests.
A smarter way to ask for help
For students, a sensible starting point is to separate practice from checking. Try the problem independently, write down the reasoning and then use AI to challenge the result. If the model disagrees, ask it to explain where the method diverges rather than simply supplying a corrected final answer. Once an explanation makes sense, attempt a similar question without assistance. That keeps the student responsible for demonstrating understanding.
A teacher can apply the same distinction when setting work. If the objective is research, drafting or comparison, using AI may be part of the exercise. If the objective is to learn a method, the assessment needs some opportunity to show unaided reasoning. LiveAIWire has examined AI tutors that ask questions instead of simply handing over solutions. That design is relevant because a question can invite thought where an immediate answer might end it.
At work, the decision will depend on the stakes. A bookkeeper might reasonably ask software to categorise routine transactions, but should still understand why exceptions matter. A developer may ask AI to generate code, yet needs enough familiarity to spot security problems or explain the behaviour to colleagues. An employee who becomes entirely dependent on an assistant could face difficulty when the system is unavailable or gives a subtly wrong result.
The same principle applies to personal decisions. Let AI organise possible options or help clarify a complicated explanation, but take time to decide what matters independently. It is useful to distinguish a question intended to save time from one intended to strengthen judgement. A perfectly efficient answer is not always the most useful intervention when the goal is to become more capable.
The result is a design challenge, not a ban on AI
Developers have a reason to rethink what a successful assistant should do. If every interaction is rewarded for speed and completeness, the product may become excellent at finishing tasks while giving users fewer opportunities to practise. A system designed for learning could offer adjustable help, detect when someone is struggling productively and provide a nudge rather than taking over. Such features would need their own evaluations rather than relying on attractive educational claims.
There are broader consequences for work and education policy. Organisations adopting AI should decide which abilities they still expect people to retain. Emergency procedures, critical checking and professional judgement cannot always be delegated to a chatbot. Teams may need occasional unaided exercises or deliberate review of how staff reach conclusions, just as they might test whether an important process works during a network outage.
LiveAIWire has also explored the tension between productivity tools and sustained attention. Instant assistance can feel like a gift of time, but a faster interaction may alter the habits surrounding the work. The Berkeley-linked experiments give that concern a specific, measurable form without claiming that technology must be rejected.
The useful conclusion is practical. AI is powerful when someone needs an answer, a summary or a quick route through repetitive work. It may be less helpful when the real purpose is learning to solve the next problem alone. The ability to keep trying after the assistant disappears is a skill worth protecting, and it will not always be visible in a polished AI-assisted result.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.