Most of us don't ask whether we can trust a colleague's first draft. We read it, decide if it's good, and fix what isn't.
With AI, the same question suddenly gets big. Can you trust it? Yes or no, and the conversation goes around in circles. We talk about it as if trust is one dial, something you turn up if you like these tools and down if you don't. I get asked this a lot. Over time I've come to think that one question is hiding three, and each of them needs a different kind of check.
Can I just look at it and know?
Say I ask an agent to clean up my notes or draft an email. Is it good? Is it what I meant? Did it save me time, or did I end up rewriting half of it? Those are real questions, but trust barely comes into it. I read the result, compare it to what I had in mind, and decide. Some of it is just taste. It's the same as reading a colleague's first draft, and most people wouldn't call that a trust problem.
This only works when I know what I wanted, though. Try asking an agent for a game that's fun to play, and nothing else. It builds one, you play it, and then what? Was it done well? I can't honestly tell you, because "fun" was never defined. Nothing went wrong. A vague ask just met a very specific answer. Hand a colleague "make it pop" and you'd get the same thing back.
I think this explains a lot of the "it just isn't that good" verdicts. One person gives an agent a single line and gets something generic. Another pours in everything they know, unfiltered, and gets something that fits. Then they compare notes on "the output" as if the tool were the only thing that changed. It's the difference between briefing a new colleague in one sentence and briefing them for half an hour. The judging is still yours, but it only works when the ask carried what you meant.
Can I make it show its work?
Some things I can't judge by looking, because I don't already know the answer. A number, a date, how something works. My own taste is no help here, so the check changes. I ask for sources. I ask it to check current information instead of going on what it remembers. It's the same move as asking a colleague where the number in their report came from. Most people don't take that personally.
Can it do damage before I ever see it?
This is where the first two checks stop working. Push to main. Send the email. Delete the file. If something can act on its own, looking at the result afterward is too late, and asking it nicely not to isn't much of a safeguard either.
You wouldn't hand a new hire a master key. Some of those doors lead to rooms where they could get hurt, or break something expensive, without ever meaning to. So you give them a keycard that opens what they need and nothing else. That says nothing about whether you trust them. It's just good building design, whoever the person is. For an agent, the equivalent is a rule the tool itself enforces, one that blocks the action outright no matter what the agent decides in the moment.
That closes off one category, things that can't be undone, and it does nothing for the other two.
So, one dial?
I would say a lot of arguments about AI are two people answering different questions. Someone who has been handed a confident, made-up fact is thinking about the second one. Someone who lets an agent tidy their notes every morning is thinking about the first. Someone who has watched an agent delete the wrong thing is thinking about the third. They're each right about their own question.
So I don't think anyone has to land on "trust it" or "don't." What works, at least for me, is matching the check to the job. I look at it when I can judge it myself. I ask for the source when I can't. And only when the stakes really call for it do I take the decision out of its hands entirely.
It's the same bar most of us already use with people. We just don't usually notice we're doing it.