Why does who can override an AI matter more than whether its advice is actually good?

This explores why control over when and how AI gets used can matter more than whether the AI gives good advice. One caveat first: none of the retrieved notes contains a study that directly compares deployment discretion against advice quality and shows discretion winning. What the corpus does have is a set of findings from different settings that all point the same way. Better advice doesn't decide outcomes on its own. Who holds the decision does.

The clearest evidence comes from studies where adequate output still wasn't enough. In one study of students handing tasks to an AI agent, trust dropped sharply on tasks that were irreversible and visible to other people, like sending an email. That happened even when people rated the agent's work as fine. High-stakes tasks that could be corrected didn't trigger the same reaction What makes people distrust AI agents they delegate to?. So people weren't reacting to the quality of the advice. They were reacting to losing the ability to step in. A study of junior and senior engineers found something similar one level up: company rules about tools, allow-lists and data policies set how much control engineers kept over agentic AI before personal preference or tool quality came into play Does personal preference shape how engineers use AI tools?. Discretion is often decided before the advice ever shows up.

The design research suggests why discretion pays off in practice. An AI research system did best when it sent only its high-uncertainty decisions to a human. That mode had an 87.5% accept rate, against 25% for full autonomy and 50% for step-by-step review Does targeted human oversight beat both full autonomy and exhaustive review?. The AI was the same in all three modes. The only thing that changed was who decided, and when. The 'Learning to Guide' approach goes further: instead of having the machine make decisions, it has the machine point out what the human should notice. That reduces anchoring on the AI's answer and keeps responsibility with the person Can AI guidance reduce anchoring bias better than AI decisions?. A philosophical account reaches a similar place. AI output should count as one piece of evidence among others, and people should stop deferring to it when it's outside its domain, biased, or contradicted by new evidence Should AI outputs replace or supplement human judgment?. Those judgments only work if someone keeps the discretion to make them.

Taken together, this suggests advice quality is necessary but rarely the bottleneck. A historical look at AI agents, from GPS onward, argues that deployments stall when conditions like trustworthiness, social acceptability and fit are missing, not when capability falls short Why do capable AI agents still fail in real deployments?. Discretion also has a social side. In four experiments, people who used AI expected colleagues to see them as less competent, and they were less willing to tell managers about it Do people fear judgment when they use AI at work?. If people can't choose openly when to use AI, they may hide good advice or use it badly.

The thread connecting these notes is that the useful thing to design often isn't a better answer. It's a better arrangement of who can override, undo, or decline the AI, and at which moments.