Kush_EXE

Ten Minutes of ChatGPT Made People Quit Sooner, Study Finds

· Runs 2:43

Researchers gave people about ten minutes of AI help on fraction or reading problems, then took it away. In three randomized trials with 1,222 people, those who had the AI got more of the final problems wrong, and in two of the trials they skipped more. In the fraction study, people who had AI solved 57% of the last three problems; people who never had it solved 73%.

The useful part is how people used it. In the bigger repeat, people who mostly asked for straight answers did worse afterward, while people who asked for hints or explanations did about as well as people who never had AI. The authors say that split is a correlation, not proof, and the sessions were short and low-stakes. Here is what the peer-reviewed paper found, what it doesn't prove, and how to use AI without losing your own skills.

Watch on YouTube Subscribe

In this video

  1. Ten minutes of ChatGPT, then it's gone
  2. Put yourself in the experiment
  3. The numbers: more wrong answers, more skips
  4. Answers vs hints: how people used it
  5. The limits of this study
  6. How to use AI and keep your skills

Ten minutes of ChatGPT, then it's gone 0:00

You ask ChatGPT for help for about ten minutes. Then it's taken away. Suddenly you get more wrong, and you're more likely to give up. An early draft made headlines in the spring. Now the peer-reviewed version is out: three randomized trials, 1,222 people.

Put yourself in the experiment 0:18

So let's put you in the experiment. You're paid a few bucks online to solve fifteen fraction problems. By random draw, you get ChatGPT in a sidebar. It already knows every answer, and hands it over whenever you ask. For the first twelve problems, you're cruising. You're actually solving more than the people without AI. Then the AI disappears. Three problems left. Just you. Almost immediately, you start getting them wrong. Or you just hit skip.

The numbers: more wrong answers, more skips 0:46

Overall, people who'd had AI solved 57 percent of those last three. People who never had it? 73. They also skipped about one in five, nearly double the no-AI group. Co-author Rachit Dubey put it this way: people weren't just giving wrong answers. They weren't willing to try without AI. Same with SAT-style reading: 76 percent solved versus 89, and eight percent skipped versus one. A bigger repeat with about 600 people found a smaller gap, 71 versus 77 percent, and no clear difference in skipping.

Answers vs hints: how people used it 1:19

Now here's the part that matters for you. That bigger trial asked people how they'd used the AI. About six in ten mostly asked it for straight answers. Afterward, they solved 65 percent and skipped 13. About a quarter asked for hints or explanations instead. They solved 76 percent and skipped 5, close to people who never had AI. Careful, though. People chose how to use it, so that's a correlation, and the authors say it's not necessarily cause and effect.

A school trial in Turkey points the same way: students using plain GPT-4 did worse on the exam. A hint-giving tutor version didn't.

The limits of this study 1:56

And the limits are real. Sessions were short, people were tested right away, and the AI only gave full answers. Stakes were low, too: a free skip button, no penalty for wrong answers. And nobody knows if it fades or builds up over time. Co-author Brian Christian thinks it's bigger than homework: "This is not a story about fractions and SAT problems."

How to use AI and keep your skills 2:16

Here's how to use AI and keep your skills. Try it yourself first. Stuck? Ask for a hint, not the answer. ChatGPT's Study Mode and Gemini's Guided Learning are built to coach like that. Just know nobody's tested them for this yet. My verdict? This doesn't prove AI makes you dumber. But ready-made answers can make you quit sooner. Make AI your coach, not your answer key.

What the sources say

  • The 2026 study "AI Assistance Reduces Persistence and Hurts Independent Performance" reports three randomized controlled trials that recruited 1,222 people in total: 354 and 667 for two experiments on fraction problems and 201 for one on SAT-style reading comprehension. arXiv, UC Berkeley News
  • The AI assistant in the 2026 study on AI assistance and persistence was GPT-5, shown to participants as ChatGPT in a sidebar. It had been given each problem and its solution in advance, so it gave complete answers on demand. arXiv
  • In the first experiment of the 2026 study on AI assistance and persistence (307 people analyzed), participants who had an AI assistant for 12 fraction problems then solved 57% of 3 unassisted test problems, against 73% for the control group, and skipped 20% of them, against 11%. arXiv
  • In the third experiment of the 2026 study on AI assistance and persistence, on SAT-style reading comprehension (168 people analyzed), participants who had used the AI assistant solved 76% of the test problems, against 89% for the control group, and skipped 8%, against 1%. arXiv
  • The second experiment of the 2026 study on AI assistance and persistence repeated the fraction test with 585 people analyzed. The AI group solved 71% of the test problems, against 77% for the control group; the gap in skipped problems, 10% against 7%, was not statistically significant. arXiv
  • In the second experiment of the 2026 study on AI assistance and persistence, 61% of the AI group said they mainly used it for direct answers and 27% for hints or clarification. The first group solved 65% of test problems and skipped 13%; the second, 76% and 5%. The authors call the findings "not necessarily causal". arXiv
  • The authors of the 2026 study on AI assistance and persistence list its limits: only brief exposure to AI was tested; people were tested immediately after the AI was removed, so how long the effect lasts is unknown; and the AI gave complete answers on demand, so whether a tutor-style AI avoids the cost is untested. arXiv
  • UC Berkeley News reported on 9 October 2026 that the updated, peer-reviewed version of the paper "AI Assistance Reduces Persistence and Hurts Independent Performance" was presented that week at the Conference on Language Modeling (COLM). The paper is on the COLM 2026 list of accepted papers. UC Berkeley News, Conference on Language Modeling
  • Knowledge at Wharton reported a trial with about 1,000 Turkish high-school students: those who practiced with a standard GPT-4 interface scored 17% worse on the exam than the control group once the AI was removed, while a group with a hint-giving GPT tutor scored about the same as the control group. Knowledge at Wharton
  • ChatGPT's Study Mode, launched on 29 July 2025, uses questions and hints instead of answers, and Google's Guided Learning in Gemini followed on 6 August 2025. The Decoder reported that no results on Study Mode's effect on learning had been published when it launched. The Decoder, Google

Sources

  1. Using AI for just 10 minutes erodes your ability to persist at hard things UC Berkeley News
  2. AI Assistance Reduces Persistence and Hurts Independent Performance arXiv
  3. AI Assistance Reduces Persistence and Hurts Independent Performance arXiv
  4. COLM 2026 Accepted Papers Conference on Language Modeling
  5. There's yet another study about how bad AI is for our brains Engadget
  6. Using AI for just 10 minutes can weaken cognitive skills, study warns KTLA
  7. Without Guardrails, Generative AI Can Harm Education Knowledge at Wharton
  8. OpenAI launches Study Mode for ChatGPT The Decoder
  9. Guided Learning in Gemini Google

More in AI news explained