Earning Trust
Trust isn't a feature you toggle on. It's a UX, a process and a contract.
Trust is a UX problem
Engineers love to treat trust as a correctness problem. "If the model is right enough, people will trust it."
That isn't how humans work. People trust systems that are:
- Predictable, same input, same shape of answer.
- Legible, I can see what it did and why.
- Recoverable, when it's wrong, I can fix it without drama.
- Bounded, I know what it won't do.
A highly accurate model that nobody trusts is less useful than a slightly less accurate one that an entire team relies on. The difference is almost always one of those four things.
The trust contract
Treat every AI feature as a contract with the user:
"I will tell you what I'm about to do, do it inside these bounds, show you what I did, and let you undo it."
If your feature can't honour that sentence, the answer isn't "ship anyway and add a disclaimer." The answer is "go back to design."
A concrete pattern
For agentic actions in GitHub Copilot, the trust contract often looks like:
- Plan, show the steps before running.
- Confirm, let the human approve, edit or reject.
- Execute, stream the work, file by file.
- Diff, show what changed, in a reviewable form.
- Revert, one button, no questions.
That loop is why people trust agent mode. Take any of those steps out and trust collapses.
