Waiting for better tools is deferred learning

The most interesting thing in Anthropic's latest usage data is not that experienced users succeed more often. It is what they stop doing: delegating. In the report Learning curves, built on a million conversations from February 2026, users with long experience iterate more, use Claude more in their work and hand less responsibility to the model in sweeping assignments. Experience does not move towards more autopilot, then, but towards less.

The beginner treats the model as an expert: asks the question, takes the answer. The experienced user treats it as a promising intern. The intern gets tasks, gets follow-up questions, gets praise when it is earned, and never gets the last word. This is not distrust. It is a division of labour someone has arrived at by trying things out.

The numbers point the same way. Users with at least six months of experience succeed more often in their conversations, and the advantage survives the controls: four percentage points remain when task type, model, use case, country and language are held constant. The report calls the result consistent with learning by doing and draws nothing more from it. The measure is also Claude's own assessment of whether the conversation succeeded, and the report itself points to alternative explanations: those who keep using the tool may be the ones whose tasks suit it, and early users may be more technical than later ones. Correlation, not proven cause.

An experiment using a completely different method points the same way, but only if you read it all the way through. In a separate Anthropic study, 52 developers were set to learn a new code library, one group with AI help and one without. The headline result goes against the tool: the AI group scored 50 percent on the knowledge test afterwards, the hand coders 67. Roughly two grades apart, with the largest gap on the debugging questions.

Inside the AI group the results spread out. The ones who did best were the ones who used the model actively: asked for explanations, asked conceptual questions, debugged on their own. They ended up level with the hand coders. The ones who simply let it generate finished fastest and learned the least. Neither source refers to the other, but they describe the same behaviour. What experienced users do more of in the usage data is exactly what kept the loss of knowledge down in the experiment.

This rubs against how AI tools are described: no threshold, just start typing. It is true that the tool answers anyone. But getting good work out of it consistently looks like something you practise, and if so, "wait until the tools get better" is not a neutral pause. The tools do get better every month. The learning curve still starts only on the day the use does.

If the advantage comes from practice, the question is not whether the tool is good enough. The question is how long your own curve is on the day you start.