Why do people hand good decisions to machines on purpose, even knowing they'd judge those calls better themselves?

This explores whether good decisions go unmade because judgment is a limited resource: people can only give careful attention to so many things. The corpus says yes, but with a twist. The best decision may be skipped on purpose, as a sensible way to spend scarce attention. Agrawal, Gans and Goldfarb show that managers will hand decisions to machines even when their own judgment would do better on that particular decision Do managers delegate to machines even when human judgment is better?. Judgment takes deliberate effort, so you have to choose where to spend it. Settling for a worse call on some decisions frees attention for the ones that matter most. Seen this way, a good decision left unmade isn't always a failure. Sometimes it's a fair trade.

The trade gets harder when AI makes the supply side cheap. If machines can produce analyses, drafts and options almost for free, the bottleneck moves to the person who has to evaluate them and own the result. One note argues that accountable judgment (verifying outputs, accepting the consequences, learning from practice) becomes the scarce asset in an economy of cheap but fallible cognition What makes accountable judgment scarce when AI cognition is cheap?. A related note takes this to its extreme with "epistemic hyperinflation": AI produces claims faster than humans can check them, so trust in knowledge loses value the way money does under inflation Can AI generate knowledge faster than humans can evaluate it?. In this picture, limited judgment doesn't just leave some good decisions unmade. It means more and more of what we decide rests on things nobody actually evaluated.

One lateral surprise is that AI systems have their own version of this problem, and engineers solve it the way the economists describe. Agents can't afford to deliberate on every step either. The SAND approach samples several candidate actions and only stops to think hard when those samples disagree When should an agent actually stop and deliberate?. The reason isn't only cost: more thinking isn't automatically better. Accuracy on one benchmark fell from 87% to 70% as thinking length grew, because models overthink easy problems and underthink hard ones Does more thinking time always improve reasoning accuracy?. ReBalance uses the model's own confidence signals to steer between those two failure modes Can confidence patterns reveal overthinking versus underthinking?. So the real skill isn't having more judgment. It's knowing where to spend it.

There's an important caveat about how far the capacity story goes. LLM agents reproduce human-like choice biases even though they have none of the attention or memory limits that are supposed to cause those biases in people Where do LLM choice biases actually come from?. That suggests some bad or missing decisions come from learned habits and incentives, not from running out of capacity. If you gave a person unlimited attention, they might still decide badly for those other reasons. Limited judgment explains part of why good decisions go unmade, but not all of it.

The corpus is strongest on the economics of delegation and on how AI systems allocate their own reasoning. It has little direct evidence on human organizations, such as studies measuring which good decisions actually get dropped under attention load. For that side of the question, the Agrawal, Gans and Goldfarb note is the best doorway.