Ten minutes with a chatbot and people give up faster, a new study finds

Researchers from UC Berkeley, Carnegie Mellon, MIT, Oxford and UCLA ran a simple experiment. They gave people fraction problems. Some had ChatGPT open in a side panel. Partway through, the researchers took it away.

The AI group started out ahead, answering more accurately than the people working alone. After problem 12 of 15 the chatbot vanished, and their accuracy dropped hard while the control group held steady. They also started skipping questions far more often. They hadn't only been leaning on the tool for answers. They'd lost some of their willingness to grind.

The paper, "AI Assistance Reduces Persistence and Hurts Independent Performance," covers three randomized experiments with 1,222 online participants: 354 and 667 on fractions, and 201 on SAT reading comprehension. It was presented at the Conference on Language Modeling. The "ten minutes" in the headlines is roughly how long people used the assistant before it was pulled.

"You almost couldn't ask for a clearer story"

That's co-author Brian Christian, a research fellow at Berkeley's Center for Human-Compatible AI and the author of The Alignment Problem. He was quick to add that fractions and SAT questions were never the point. The study is a prompt to rethink what "helpful" should mean for software that now sits in every school and office.

The usual caveats apply. The tasks were short, narrow and run online, and the study doesn't show what happens over months of real work. Anyone who has reached for autocomplete before finishing their own thought will still recognize the effect. The tool helps right up until it's gone, and then you find out how much of the work it was doing.

Sources: EurekAlert, TechXplore, News-Medical.