Does AI really save time, or just change how we spend it?

A common assumption about AI productivity is that it reduces the time required to complete a task. The mechanism is straightforward: AI does some of the work, the worker does less work, total time goes down. The Skill Formation study finds the picture is more complex. AI does not necessarily reduce total task time — it shifts how the time is spent.

In the study, participants using AI showed less active coding time than the control group. But the saved coding time did not translate into saved total time. Instead, time shifted to two new categories: composing AI queries (some participants spent up to 11 minutes on this within a task) and reading/understanding AI-generated content. The total time-on-task was not significantly different from controls; the time was spent on different activities.

This reallocation matters for several reasons. First, the activities AI introduces (prompt composition, output evaluation) are different cognitive operations than the activities AI replaces (active problem-solving, coding). The worker is not simply doing less work; they are doing different work. Whether the new work is more or less cognitively valuable depends on context. Second, the new work often produces no durable artifact. Time spent composing a query that produces output the worker uses and discards leaves no trace; time spent solving a problem produces a solution-skill the worker retains. Third, the learning outcomes track the activities, not the total time. Workers who spent time understanding AI generations (rather than only generating with them) learned more — the activity, not the AI use itself, drove the learning effect.

The diagnostic implication for design and management: time-on-task is a poor proxy for AI value. Two workers may complete the same task in the same time, one having learned much and one having learned nothing. The difference is in what they spent the time on, not how long they took. Productivity metrics that ignore this conflate activities with very different downstream consequences.

For the worker, the implication is that AI introduces a new category of attention-cost: the cost of evaluating AI output, deciding what to do with it, and integrating it. This cost is invisible in the standard productivity measurements but is real and substantial. The productivity gain depends on whether the new attention-cost is less than the old work-cost, which is not guaranteed.

The strongest counterargument: better interfaces (voice, agentic, ambient) will reduce the AI-interaction overhead. Possible, but each such interface introduces its own attention-pattern; the cost-shift is to a different kind of attention rather than to no attention at all.