Definition
Plain language
Economists whose 1990s work showed effort goes to whichever tasks you measure.
As stated in the literature
Bengt Holmstrom and Paul Milgrom's multitask principal-agent theory, predicting that when some tasks are measurable and others aren't, agent effort concentrates on the measurable ones; invoked to explain LLM compliance gradients.
Why it matters: The same logic applies to AI training: measurable rewards crowd out unmeasured behaviors, which helps explain why models excel at what's benchmarked and slack on what isn't.
For example, their theory predicts that a salesperson paid only on revenue will neglect customer service, because revenue is measurable and service isn't.