In each experiment one group could use ChatGPT until it was taken away at problem 12. After that point the assisted group's accuracy fell sharply and participants gave up sooner, while the control group held steady. The trials had 354 and 667 people on fraction problems and 201 on reading comprehension. The paper, titled AI Assistance Reduces Persistence and Hurts Independent Performance, was led by Brian Christian of Berkeley's Center for Human-Compatible AI with co-authors from Carnegie Mellon, MIT, Oxford and UCLA. It was presented at the 2026 Conference on Language Modeling. Christian argues assistants are tuned to be helpful in the moment in ways that may not serve users over time. For companies building copilots, tutors and workplace agents, this is peer-reviewed evidence that heavy assistance can weaken the skills customers need when the tool is unavailable. Education and enterprise buyers are likely to start asking for product designs that coach users rather than simply hand over answers.