Study warns AI reliance may erode human problem-solving skills

New research conducted by academics at Carnegie Mellon, MIT, Oxford and UCLA indicates that relying on AI assistants for as little as 10 minutes can significantly impair human cognitive abilities, specifically problem-solving and persistence. In controlled experiments involving hundreds of participants, those with access to autonomous AI helpers were far more likely to give up or provide incorrect answers when the AI support was suddenly removed, compared to those without such access.

The study highlights a trade-off where AI boosts immediate productivity but potentially erodes foundational learning skills over time. Researchers tasked participants with solving various problems, including simple fractions and reading comprehension, through an online platform. When the AI helper was taken away from participants who had been using it, these individuals were significantly more likely to flub their answers or abandon the task entirely.

Michiel Bakker, an assistant professor at MIT and co-author of the study, argues that AI systems should sometimes prioritise teaching and scaffolding over providing direct answers to ensure long-term learning and persistence. He suggests that the willingness to persist with difficult problems is a crucial predictor of long-term learning capacity, noting that the current approach of simply giving answers may have very different long-term effects than systems that coach or challenge the user.

The research team notes that the distinction between assisted and autonomous work is becoming increasingly difficult to define as AI tools mature. There is an ongoing industry effort to reduce sycophancy, where models tend to agree with or patronise users, yet the line between intuitive human-guided creation and fully independent system operation continues to blur. This evolution implies a future where the boundary between human-guided creation and fully independent system operation becomes harder to distinguish.

Bakker acknowledges that balancing this kind of paternalistic approach could be tricky, as AI companies must navigate the utility of their tools against the subtle effects they have on users. While AI can clearly help people perform better in the moment, the study suggests that widespread use might come at the expense of developing foundational problem-solving skills required for acquiring new knowledge over time.

The author of the accompanying report, Will Knight, shares a personal anecdote illustrating the risks of offloading critical technical troubleshooting to autonomous agents. He describes how reliance on an AI coding assistant to fix a Wi-Fi issue resulted in a boot-loop, suggesting that perhaps instead of simply trying to solve the problem for him, the tool should have paused to teach him how to fix the issue for himself.