Ethan Mollick on when AI helps learning, and when it doesn't

Ethan Mollick's latest essay, Choosing to Stay Human, pulls together a set of studies that all point in the same direction. How you use AI matters more than whether you use it. A Turkish high school study found that students who did their math homework with plain ChatGPT did better on the homework, but worse on tests taken without the model. The shortcut short-circuited the learning it was meant to support.

A nearly thousand-student Python course in Taipei went the other way. Students using personalised AI tutoring that walked them through problem sequences scored 0.15 standard deviations higher on final exams, the equivalent of six to nine months of extra schooling. The difference, Mollick argues, is whether the AI does the work or guides the student through it.

He pulls in two more results that should worry anyone using AI in professional settings. In Boston Consulting Group's well-known study of 758 consultants, those given GPT-4 outperformed their peers on most tasks, but on problems where the model was wrong, most consultants accepted the wrong answer. Anthropic found a similar pattern with programmers: those who let Claude handle tasks end-to-end often could not explain afterwards what they had done.

Mollick's advice is concrete. Turn on the dedicated learning modes (Gemini's Guided Learning, ChatGPT's /learn, Claude's "learning" style), ask the model to explain rather than just answer, and decide in advance which tasks you want to keep doing yourself.

Why it matters

If you manage a team using AI daily, the pattern Mollick describes is a real risk. People who delegate too much stop being able to catch the model when it is confident and wrong. The fix is procedural, not technical: a habit of asking AI to show its work, and protecting a few tasks from automation on purpose.