Schneier gives AI delegation a gym-task test

Bruce Schneier split AI-delegation tasks into two categories, in a post Simon Willison surfaced on July 30. A task can be one a model handles perfectly. It can still be wrong to hand off, if doing it yourself is what teaches you to catch the model's mistakes later.

Schneier's test asks whether a task's value lives in the artifact it produces, or in the capability doing it builds in you. He calls the second kind a gym task. "The writing assignments I give my students are gym tasks, not work tasks," he wrote in the post. Skip the rep, and the skill it was building never forms.

Most AI-delegation arguments fixate on capability: can the model do this well enough to trust? Schneier's frame sidesteps that fight entirely.

I hand a coding agent the boilerplate and the routine migrations without a second thought. Those are work tasks. The code is the point. I still write the gnarly concurrency logic myself, because that's the rep that keeps me able to tell when the agent's version is wrong.

The rule outlasts the usual capability debate because it doesn't depend on the model staying weaker than you. Even once AI writes better code than you do, the writing is still what builds the judgment to check it. Watch which engineers stop being able to spot bad AI output first. They'll be the ones who delegated their gym tasks.