← Back to Blog
AI & Automationby ELHASSANE Studio

Ten Minutes With AI Can Weaken Your Persistence

Ten Minutes With AI Can Weaken Your Persistence

A chatbot answers in seconds. A hard problem used to take an hour. Something shifts in how the second thing starts to feel — and now there is a paper that measures the shift instead of only describing it.

In October 2026, UC Berkeley’s research office summarized a study that had just been presented at the Conference on Language Modeling, “AI Assistance Reduces Persistence and Hurts Independent Performance.” The team — Grace Liu (Carnegie Mellon), Brian Christian and Tsvetomira Dumbalska (Oxford), Michiel Bakker (MIT) and Rachit Dubey (UCLA) — ran three randomized controlled trials with 1,222 participants. The result is blunt: AI help made people better in the moment, and worse the moment the help disappeared.

The experiment in three parts

Experiment 1. A total of 354 participants took 15 fraction problems. Half had a GPT-5 sidebar pre-loaded with each problem and its solution; half worked alone. After 12 problems, the AI was removed without warning, and everyone solved the same 3 final problems. On those problems the AI group solved 0.57 of them, against 0.73 for the group that never used AI. They also gave up more often: a skip rate of 0.20 against 0.11. Solving fell, and quitting rose.

Experiment 2. A total of 667 participants repeated the test with two fixes: a pretest to screen for real skill, and a matching sidebar for the control group so that both groups saw the same interface disappear. The pattern held. Without the AI, the AI group solved 0.71 against 0.77 for the control group. The skip-rate gap this time — 0.10 against 0.07 — was in the same direction but did not reach statistical significance. Then came the useful split. Of the AI group, 61% said they used it to get answers directly, 27% for hints or clarification, and 12% said they never used it. Only the first group was damaged: those who took direct answers solved 0.65 without the tool, and finished below their own starting level. Hints did not carry that cost. The variable is not “used AI.” It is “let AI do the thinking.”

Experiment 3. A total of 201 participants moved from fractions to SAT reading comprehension. Same design, same outcome. Once the tool was gone, the AI group solved 0.76 against 0.89, and skipped 0.08 of the questions against 0.01. The effect is not a quirk of arithmetic; it survives on a task built on reading and reasoning.

The finding has an asterisk

This is a causal, short-term result, and the authors are careful about how far it travels. They “urge caution in interpreting these results as broad evidence that AI lowers persistence and independent ability for all tasks.” The long-term, cumulative damage is stated as a hypothesis — what might happen “if such effects accumulate” over months and years — not something this study measured. Read honestly, the strongest claim here is narrower and more useful: a short, unremarkable session of offloading changes how you behave when the tool is gone. In Experiment 2, even the persistence signal itself was only directional. The headline is bigger than the effect size, and the paper says so.

Why it could happen

The team proposes two mechanisms. The first is a shift in expectation: when AI completes work in seconds, the benchmark for how long a task should take moves, and unaided effort starts to feel heavier than before. Each act of offloading raises the cost of the next unaided attempt, which makes the next offload more tempting. The second is the disappearance of what the researchers call “productive struggle” — the part of work where you not only learn the material but learn what you are capable of without help. Remove it, and both knowledge and self-knowledge thin out.

What to do with this

  • Ask for hints, not answers. The clearest line in the data separates people who used AI to get answers from people who used it to understand. Only the first group paid.
  • Keep one block of work that AI never touches. Twenty minutes a day on something genuinely hard, with no sidebar, is cheap insurance.
  • Notice the withdrawal test. The measure in every experiment was simple: take the tool away and watch. Try that on yourself this week on a real task.
  • Treat the speed as the warning. The moment “this is taking too long” starts to mean “this is wrong,” the expectation has already moved.

None of this is an argument against AI. It is an argument against deploying it as an oracle. The systems are extraordinarily helpful in the instant and indifferent to who you become through using them — and the difference between the two is the design choice we keep making by habit.

If the productivity angle is what brought you here, read why AI can boom while productivity stalls. And if you want a slow, hands-on counterweight, coloring is the cheapest one we know.

Sources

#ai#productivity#trending