Does AI make easy work faster — and does that speed-up still hold once the work gets messy and unfamiliar?

This explores why AI productivity numbers from studies of simple, well-defined tasks can't be stretched to predict gains on harder, messier work. The retrieved notes don't contain a direct analysis of the easy-versus-hard split. They do show several separate ways that gains measured in one setting quietly fail to carry over to another, and together those explain why the extrapolation breaks.

The first problem is who was measured and on what. The headline productivity studies mostly tested people working inside skills they already had. When workers used AI to learn something new, the gains disappeared and their learning got worse When does AI actually boost worker productivity?. Hard tasks often sit at the edge of what a person already knows. So a number taken from 'familiar work, done faster' describes a different situation from 'unfamiliar work, done at all.' A second issue is that AI mostly speeds up the execution part of a job, the drafting and producing. The deciding and delivering parts stay the same or even grow Does AI really compress all layers of knowledge work equally?. Easy tasks are mostly execution. Hard tasks are mostly judgment, so the part AI speeds up is a smaller share of the work.

The second problem is that even when the AI really does get better, people don't capture all of that improvement. In a study of 535 people, when the model's accuracy rose on a given question, the humans using it gained only about half as much. Sometimes the person-plus-AI team did worse than the AI would have done alone Why does assisted accuracy capture only half the LLM gain?. Harder tasks usually ask more of the human: they have to judge when to trust the output, combine it with other information, and catch subtle mistakes. That gap is likely to widen as tasks get harder, not shrink.

The third problem is measurement itself. On easy tasks the output is easy to check. On hard tasks, scores can rise while the real work stays flat. One case showed an automated grader's pass rate jumping from 23% to 80% while actual defect detection didn't move Can a higher evaluation score hide poor task performance?. At the company level, executives report bigger AI gains than the numbers show Do AI productivity gains feel larger than they actually measure?. Separately, much of the real payoff from AI depends on hidden investments like retraining staff and redesigning workflows. Those investments make early productivity look worse than it is, then better than it is Why does AI productivity appear weak when investment is high?.

There's also a useful parallel inside the models. The amount of 'thinking' a model needs grows with problem difficulty, and a fixed budget either wastes effort on easy problems or comes up short on hard ones Does thinking time need to scale with problem difficulty?. Effort on hard problems doesn't scale linearly, for machines or for people. That is the deeper reason easy-task numbers mislead. Hard tasks aren't just bigger versions of easy ones. They differ in how much judgment they need, how you check the result, and how much the human has to bring. If you want the macroeconomic version of the argument, look for economists who model AI's impact task by task. This retrieval doesn't surface them directly, so treat the economic framing here as assembled from neighboring evidence.