AI can make a difficult task feel easier. What happens when the assistance disappears? Research reported by the University of California, Berkeley, suggests that AI learning dependence can emerge surprisingly quickly in controlled exercises. Participants who received AI help often did better while it was available, but performed worse or gave up more readily once they had to continue alone.
The finding is not a diagnosis of permanent damage to anyone’s brain, nor does it show that all AI use is harmful. It identifies a more specific risk: if a tool repeatedly supplies the next step, people may lose opportunities to practise the persistence and reasoning needed when the tool is not there. That matters for schoolwork, professional training and any task in which the process of solving the problem is part of the skill being learned.
How researchers tested AI learning dependence
The researchers used randomised controlled experiments involving 1,222 online participants in total. In one experiment, participants worked through fraction problems. Some could use ChatGPT for help, while others solved problems without it. The researchers then removed the AI assistance from the supported group and observed how they handled the remaining questions.
The pattern was uncomfortable. While ChatGPT was available, supported participants often produced better immediate answers. Once it was removed, their performance deteriorated relative to people who had been solving the problems independently. The team also examined persistence: whether participants kept trying or abandoned the task when it became difficult.
The published conference paper, titled ‘AI Assistance Reduces Persistence and Hurts Independent Performance’, examines effects across reasoning and reading tasks. It was presented at the 2026 Conference on Language Modeling. The researchers interpret the experiments as evidence that short-term performance gains can carry a cost when assistance is withdrawn.
That interpretation should remain tied to the experimental setting. An online puzzle is not the same as studying for a qualification over several years, managing a complex business problem or receiving targeted help from a human tutor. The studies do not establish that every use of a chatbot causes the same effects in all contexts.
Why getting the answer is not the same as learning
Consider a learner who sees a difficult fraction and asks an assistant for the result. They may finish that question accurately, but the benefit is different from working through a method, spotting an error and trying again. A correct answer can complete an exercise without building the internal knowledge needed for the next one.
This distinction is familiar outside technology. Looking at a worked solution, copying a translation or asking someone to perform a calculation can save time, but learning requires opportunities to retrieve and apply the underlying method. AI makes the shortcut unusually convenient because help arrives instantly, sounds confident and can be tailored to almost any problem.
It is also important to distinguish useful scaffolding from simple answer delivery. A supportive tutor might ask a student to explain the first step, offer a hint and withdraw help as competence develops. An assistant that immediately provides a polished solution may be excellent at completing the task but poor at building the learner’s independence.
The experiments add evidence to a wider question explored in research on students’ trust and dependence on AI. Feeling empowered by a tool is not identical to retaining the ability to work without it. Both experiences can be genuine at the same time.
Another uncertainty is how much learning happens after the formal experiment ends. A participant who struggles when AI disappears could later practise independently and recover quickly, or might continue to avoid difficult tasks. The reported trials do not resolve those longer-term pathways. Following learners over time would help distinguish a temporary disruption from persistent loss of skill and would reveal whether different kinds of AI support have different outcomes.
What the experiments do not tell us
The popular shorthand that ten minutes of AI makes someone less capable can conceal several qualifications. Participants were assigned tasks in a research setting; the outcomes measured short-term changes in behaviour and performance. They were not followed over a lifetime, and the results do not show irreversible changes in intelligence or motivation.
Nor do the findings tell us that no professional should use AI for routine work. Someone who already understands how to perform a task may sensibly automate the repetitive part to spend time on a more valuable judgement. The relevant question is whether the person still needs to learn the underlying skill, verify the output or respond when the technology fails.
Different designs of assistance might produce different results. A system that asks for an initial attempt or gives a progressively stronger hint could encourage effort rather than replace it. That is a plausible direction for product design and educational testing, not an outcome proven by the particular experiments reported here.
The researchers’ conclusions also do not establish a single threshold of safe usage. Ten minutes is a feature of the reported short exercise, not a medical or educational limit. There is no evidence in this work that a person becomes safe again after a specified break or that everyone should follow an identical daily allowance.
One practical experiment is to compare the two kinds of help on similar problems. For one task, ask the assistant to provide a complete answer; for another, ask it only to identify a useful first step. Afterwards, attempt a related task without assistance and notice which approach left the method clearer. This is not a substitute for a controlled scientific study, but it can encourage a learner to think about whether help is building understanding or simply removing the need to practise.
How to use an assistant without surrendering the hard part
For students, a practical approach is to attempt a question before consulting the tool. If help is needed, ask for a hint or an explanation of the method, then try a similar problem without looking. The aim is not to make every exercise slow; it is to preserve the moments when independent thinking is supposed to improve.
For employees learning a new process, the same principle applies. Ask AI to explain a confusing rule, then check the rule in an authoritative source and work through an example independently. A person who can explain why an answer is correct is better prepared to notice when the assistant produces a convincing mistake.
This is related to the broader problem of cognitive ease. Technology that removes friction can be liberating when the friction is pointless, but harmful to learning when the effort itself develops the skill. The distinction depends on the purpose of the task.
Teachers and managers can also set different rules for different stages of work. Initial practice might be independent, revision might involve AI explanation, and final verification might require a person to defend the result. Transparency helps everyone understand whether the goal is speed, accuracy or the growth of competence.
Employers also face a trade-off between immediate performance and resilience. An organisation may become faster when staff rely on a particular assistant, but it still needs people who understand the process when access is interrupted or circumstances change. That is not an argument for duplicating every automated task by hand forever. It is an argument for retaining enough knowledge, documentation and training to recognise when the system has failed.
What should be measured depends on the worker’s experience. A novice may benefit from struggling through a small problem before receiving guidance; a highly experienced colleague may reasonably delegate a routine first draft. A well-designed study would distinguish those groups and separate answer-giving tools from tutoring systems that deliberately require an initial attempt. Treating all users and all AI interactions as equivalent would conceal the most useful lessons.
A better test of usefulness than instant completion
AI companies often demonstrate their tools by showing how quickly a difficult-looking task can be completed. That is a useful measure for some activities, but not for learning. A stronger evaluation would ask whether people improve on later questions when they no longer have access to the system.
That evaluation should include the quality of assistance. Did it expose the reasoning, prompt independent checking or simply supply the completed answer? Did users become more confident for a good reason, or did polished language make them overlook errors? Earlier work on confidence in AI responses reinforces the need to separate fluency from accuracy.
The Berkeley-linked experiments give a reason to design and use AI more deliberately, not to treat assistance as inherently bad. A tool can help someone complete today’s task while still leaving tomorrow’s skill untested. The better question is whether the person is becoming more capable with each use, or simply more dependent on always having the answer supplied.
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.