AI and Trust

The conversation about artificial intelligence and trust has, so far, been almost entirely the wrong conversation. We have spent enormous energy debating whether AI is trustworthy, as though trustworthiness were a property of the model itself. It is not. Trustworthiness is a property of the entire system, the model, the people who deploy it, the institution that stands behind it, and the legal and cultural infrastructure that holds all three accountable.

A model that hallucinates inside a medical institution with strong review protocols and an honest culture of error reporting is a manageable risk. The same model deployed inside an organization that punishes the messenger and rewards confident output is a catastrophe waiting to be discovered.

Four Questions Every Board Should Ask

1. What can this system do that we cannot independently verify, and who pays the cost when it is wrong?
2. If our most cautious employee raised a concern about this deployment, would they be heard, or would they be managed?
3. What is the documented protocol for shutting it down, and has anyone actually rehearsed it?
4. Whose interests does the system optimize for when our interests and the user's interests diverge?

The Trust Stack

Trustworthy AI is a stack, not a feature. At the bottom is the model itself, its accuracy, its known failure modes, its provenance. In the middle is the deployment context, the guardrails, the human review, the data hygiene. At the top is the institution, its incentives, its culture of honest disclosure, its track record of doing the right thing when no one was forcing it to.

Every layer can fail independently. Most current AI disasters fail at the top of the stack first, and only look like model failures in the post-mortem.