Ten Minutes With AI Reduce Persistence on Hard Problems, Berkeley Study Finds

Ten minutes with an AI assistant is enough to make people worse at sticking with hard tasks, according to the University of California, Berkeley: they give up faster, try fewer approaches and look for an easy way out sooner. The piece was published on the university's research portal and collected 55 points and 65 comments on Hacker News within a few hours. The topic matters to anyone who writes code, text or analysis with the help of language models.

First, the limits. The description available to us does not include the sample size, the format of the tasks, the control groups or the size of the effect. All of that appears only in the full paper, and without it a headline remains just a headline. Below, we look at what exactly is being claimed and why it is plausible given what is already known.

What exactly was measured

The claim is not about a drop in knowledge or worse decisions. It is about persistence: the willingness to keep going when a task does not yield on the first attempt. This is a separate skill. It determines whether a person gets through a difficult bug-fixing session, a proof or a draft report.

The mechanism in such studies is usually the same. The model instantly produces an acceptable answer, and the person recalibrates their expectation of how much effort a task is worth. The next difficulty, without a prompt, feels disproportionately expensive. Ten minutes is a very short window. If the effect holds up over such a span, it is not an accumulated habit but a rapid recalibration.

What is already known on the topic

There are no direct matches to the Berkeley work, but indirect evidence is accumulating:

• In 2025, researchers at MIT Media Lab, in a study called Your Brain on ChatGPT, compared three groups of people writing essays: with an LLM, with a search engine and with no tools. According to EEG, the group using the model showed weaker connectivity in brain activity. The sample was small, about 54 people.
• A survey by Microsoft Research and Carnegie Mellon of 319 knowledge workers found a link: the more a person trusts AI, the less critical thinking they put into the task.
• Studies of automation in aviation and medicine have described a similar phenomenon for decades: an operator used to receiving prompts copes worse when they disappear.

Against this background, the Berkeley claim looks like a continuation of an existing line of research rather than a surprise. The novelty, if there is any, lies in the time scale: minutes instead of weeks.

Key parameters of the piece

• Source: UC Berkeley research portal
• Claimed threshold: 10 minutes of AI use
• Affected skill: persistence on hard tasks
• Community reaction: 55 points and 65 comments on Hacker News
• Unknown from the available description: sample, experimental design, duration of the effect

How to work with this

Until the full methodology has been examined, it is reasonable to treat the result as a working hypothesis, but protective habits are cheap. If you are a developer or analyst, first spend 15–20 minutes on your own attempt, and only then open the assistant. That way you keep the habit of tolerating uncertainty. For hard tasks, ask the model not for a ready-made answer but for guiding questions or a review of your solution. Teams should separate modes: hand routine work over to AI entirely, and leave learning and strategic tasks to humans, at least for the first pass.

Managers rolling out assistants across a whole company should track not only speed of execution but also what happens to people's independence when the tool is unavailable. This can be measured: give one hard task without AI once a quarter and compare the time to completion.

The main question for the study's authors is how long the effect lasts and whether persistence returns after a break. The answer determines whether this is an alarming signal or just a quirk of the first minutes after an easy result.