Study links direct AI answers with lower persistence when help disappears

People given AI help initially did better on maths and reading tasks, but were more likely to struggle and give up when the tools were taken away, Geo News reports. The study’s sharper distinction: participants who received direct answers showed a greater decline in persistence than those who used AI for guidance.

Watch Desk analysis

What happened

The report describes a study presented at the 2026 Conference on Language Modeling, involving 1,222 participants across three experiments. Participants either used AI while solving problems or worked without it. The AI-assisted group initially performed better and skipped fewer questions; after chatbot access was withdrawn, they struggled more with similar tasks and were more likely to give up.

The researchers, from Carnegie Mellon University, MIT, Oxford, UCLA and the University of California, Berkeley, found a greater decline in persistence among people who relied on AI for direct answers than among those who used it for guidance. Lead author Grace Liu, a Carnegie Mellon doctoral student, said the effort involved in tackling difficult problems can help develop skills. The account does not establish lasting effects: the study examined short-term outcomes and did not show permanent damage to cognitive abilities.

Why it matters

The useful question is not simply whether AI makes a task easier. It is what happens when the assistance goes away. If direct answers help people finish today’s exercise but make them less willing to tackle tomorrow’s unaided, convenience has acquired a rather awkward footnote.

That does not mean AI help is inherently harmful. The reported difference between answer-giving and guidance points to a practical design choice: a tool that nudges someone towards a solution may support learning better than one that simply supplies it. The study’s short-term scope matters, too; it cannot tell us whether these effects persist or how they translate beyond the tasks tested.

Our read

This is a useful warning against treating task completion as the whole report card. AI tutors should be judged on whether people can still think through a problem once the tutor steps aside, not only on how quickly they produce an answer. For learners, asking for a hint or an explanation rather than a finished solution is a sensible experiment, not a guarantee.

What to watch

• Whether the research paper reports more detail on the tasks and how AI assistance was delivered.
• Whether follow-up studies find lasting effects, or different results with longer-term learning.
• Whether AI learning tools test for independent problem-solving, not just immediate answers.

Discussion spark: For learning, should AI tools default to hints and guided reasoning, even when users ask for the answer, or should that choice stay entirely with the learner?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.