Reliability as Accuracy

A system is accurate if its outputs are correct within its domain. An AI model trained on medical imaging is accurate if it correctly identifies tumors in X-rays. A calculator is accurate if 2+2 equals 4. Accuracy is verifiable: you can check the output against reality.

Accuracy is valuable. It is not trust.

Here is the critical distinction: accuracy is a technical property. It can be measured, sometimes rigorously. But professional trust is not a technical property. It is a social and relational property. You can have an accurate AI system and still not trust it — because trust involves judgment about what matters, and accuracy only measures whether the system is correct within a specified domain.

Consider a medical diagnosis system that is 95% accurate at detecting a particular disease. That is impressive accuracy. But who decides whether it is ethical to use that system in a population where the base rate of the disease is very low? Who decides how to communicate the results to a patient? Who decides what to do if the system's recommendation conflicts with what the patient wants or what clinical judgment suggests? These are not accuracy questions. They are judgment questions. They are trust questions. And they require a human who can be held accountable if something goes wrong.

An AI system cannot make these choices. It cannot hold the stakes. It cannot be sued if it causes harm. It cannot be fired or demoted or promoted. It cannot be trusted in the full sense because there is no "it" to be held responsible. There is only code.