Using AI for Just 10 Minutes Might Make You Lazy and Dumb, Study Shows

Use AI chatbots Because even just 10 minutes can have an extremely negative impact on people’s ability to think and solve problems, according to a new study led by researchers from Carnegie Mellon, MIT, Oxford and UCLA.

The researchers tasked people with solving a variety of problems, including simple fractions and reading comprehension, through an online platform that paid them for their work. They conducted three experiments, each involving several hundred people. Some participants had access to an AI assistant capable of solving the problem autonomously. When the AI ​​assistant was suddenly taken away from them, these people were much more likely to abandon the problem or blunt their responses. The study suggests that widespread use of AI could boost productivity at the expense of developing fundamental problem-solving skills.

“The takeaway is not that we should ban AI in education or workplaces,” says Michiel Bakker, an MIT assistant professor involved in the study. “AI can clearly help people perform better in the moment, and that can be valuable. But we should be more careful about what kind of help AI provides and when.”

I recently met Bakker, who has messy hair and a wide smile, on the MIT campus. Originally from the Netherlands, he previously worked at Google DeepMind in London. He told me that a well-known essay about how AI could disempower humans over time inspired him to think about how technology might already be eroding people’s abilities. The essay makes for somewhat grim reading, as it suggests that loss of power is inevitable. That said, perhaps understanding how AI can help people develop their own mental abilities should be part of how models are aligned with human values.

“It’s fundamentally a cognitive question: about perseverance, about learning, and about how people respond to difficulties,” Bakker tells me. “We wanted to take into account these broader concerns about the long-term interaction between humans and AI and study them in a controlled experimental setting.”

The resulting study seems particularly concerning, Bakker says, because a person’s willingness to persist in problem solving is crucial for learning new skills and also predicts their ability to learn over time.

Bakker says it may be necessary to rethink how AI tools work so that, like a good human teacher, models sometimes prioritize learning from a person rather than solving a problem for them. “Systems that give direct answers can have very different long-term effects than those that scaffold, coach, or challenge the user,” says Bakker. He admits, however, that it could be tricky to find the right balance between this type of “paternalistic” approach.

AI companies are already thinking about the more subtle effects their models can have on users. The sycophancy of certain models – or their likelihood of agreeing with and patronizing users – is something OpenAI has sought to mitigate with new versions of GPT.

Placing too much trust in AI seems especially problematic when the tools may not behave as expected. Agentic AI systems are particularly unpredictable because they perform complex tasks independently and can introduce strange errors. One wonders what Claude Code and Codex do to the skills of coders who may sometimes need to fix the bugs they introduce.

I recently received a lesson in the danger of entrusting my own critical thinking to AI. I use OpenClaw (with Codex inside) as a daily aid and have found it remarkably effective in solving configuration problems on Linux. However, recently, after my Wi-Fi connection continued to drop, my AI assistant suggested I run a series of commands to change the driver talking to the Wi-Fi card. The result was a machine that refused to boot no matter what I did.

Maybe instead of just trying to solve the problem for me, OpenClaw should have taken a break from teaching me how to solve the problem on my own. As a result, I could have a better computer and brain.

This is an edition of Will Knight’s AI Lab Newsletter. Read previous newsletters here.