Reliability as Consistency

A tool is consistent if it behaves the same way given the same input. An AI model is consistent in this sense: feed it the same prompt, and it will produce similar outputs. A manufacturing machine is consistent: crank the handle, get the same shaped part, every time. Consistency is predictable. It is automated. It requires no judgment.

Consistency is useful. It is not trust.

Consider the difference between a colleague who is consistently competent at a narrow task and a colleague you trust with a crisis. The first is reliable in the sense of consistency. They will execute their role predictably, day after day. But when something breaks — when the situation is not the one they were trained for — they may have limited value. A colleague you trust with a crisis is something else entirely. You trust them because you believe they will figure out what the right thing is, even if they have never encountered that particular situation before. You trust their judgment, not just their consistency.

AI systems are consistent. They are not trusted in the crisis sense because they cannot bring judgment to a situation that falls outside their training domain. They cannot adapt their understanding of what matters. They cannot decide that the rule should be broken in this case, or that the stated goal is actually harmful and should be questioned.