Study: 10 minutes of AI help made people give up sooner once it was taken away

Researchers including UC Berkeley's Brian Christian, with colleagues from Carnegie Mellon, MIT, Oxford and UCLA, ran randomised trials with 1,222 people online. Berkeley says the peer-reviewed version was presented this week at the Conference on Language Modeling. In the first test, people solved basic fraction problems, and one group had ChatGPT open alongside and could even ask it for answers. That group started out more accurate. After 12 problems the AI was taken away, and their accuracy dropped almost at once while they skipped more questions. A larger second group (667 people) showed the same pattern, and a third test on SAT reading comprehension (201 people) did too. The effects appeared after about 10 minutes of use. The authors think AI trains people to expect instant answers, so they stop pushing through hard steps. These were short, online lab tasks, not real jobs. Hacker News thread: No consensus. Critics say the design is weak and should have compared AI with other help, such as a calculator, books or a human expert, and several used tool analogies (a table saw, a hammer) to argue losing a skill you no longer need is normal. Others say it matches their experience: when the AI hits a wall and you have to work it out yourself, it is noticeably harder. A few said AI has made them more productive and more persistent. Sorted take: It is a narrow study, but the risk is real for people still learning a job. Let new staff do some tasks by hand first, and set AI tools to explain the steps rather than hand over answers.