The wrong question
"Do you trust AI?" is a bad question, and I hear it constantly. It is like asking whether you trust a knife. Trust is not a property of the tool. It is a property of the tool doing a specific task in a specific situation, and the honest answer is different for each one.
So here is roughly how I calibrate it, task by task, after getting it wrong a few times in both directions, too trusting and too suspicious. The single question underneath all of it is simple: how expensive is it if this is wrong, and how easily would I catch it if it were?
Those two questions do almost all the work. Cheap to be wrong and easy to catch, let it run and barely look. Costly to be wrong and hard to catch, slow right down, because that is exactly the combination the polish is most dangerous in. Everything else sits somewhere in between, and the skill is placing each task honestly instead of defaulting to one setting for all of them.
Where I let it run
Work that is easy to verify and cheap to get wrong. Reformatting something. Drafting the boring first version of a document I am going to rewrite anyway. Summarising a thing I could check in a glance. Generating options I am going to judge, not adopt. Here the tool is fast and the downside is small, so I let it move and barely look.
This is where the tools genuinely give you your afternoon back, and where being precious about checking everything is just slow for its own sake. If catching the mistake would take ten seconds and the mistake costs nothing, let it run.
The thing to guard against here is the opposite failure, the person so burned by one bad output that they now read every reformatted list as if it were a legal contract. That is just slowness dressed up as diligence. If being wrong is cheap and catching it is trivial, spending real attention on it is not care, it is waste.
The risk isn't the AI being wrong. It's the AI being wrong in a way that looks exactly like being right.
Where I verify everything
Work where the output looks authoritative but I would carry the consequences of it being wrong. Anything going in front of a client under my name. Anything that commits us to a decision. Code touching something sensitive. Numbers I am going to act on. Here the danger is precisely that it looks finished, so I read every line as if a stranger wrote it, because in a sense one did.
This is the bucket where calibration matters most, because the output is most convincing exactly where the cost of being wrong is highest, and those two things pull against each other. The more authoritative it looks, the more tempting it is to trust, and the more it would cost you if that trust were misplaced. So this is where I slow down the most, precisely when everything about the output is telling me I do not need to.
Where I do not use it at all
The judgment calls. Whether a green status is actually true. What to say to a client who has gone quiet. Which risk is the one that matters. What we should build. These need context that only exists in the room, in the relationship, in the history, and the tool was in none of those places. Handing them to a model does not save time, it just launders a guess into something that looks considered.
The mistake I keep watching people make, including past versions of me, is trusting the tool at a single global setting: sceptical about everything and slow, or relaxed about everything and exposed. The calibration is the skill. The tool is the same. What changes is what it would cost you to be wrong.
The point of a field guide is that you do not have to decide from scratch each time. Once you have the buckets, most tasks sort themselves in a second: is this cheap to be wrong, or is it a judgment call that was never the tool's to make? Getting that sorting right, quickly, is most of what using these tools well actually is.
None of this makes the tools less useful. If anything it makes them more useful, because a tool you trust exactly as much as each task deserves is one you can lean on hard where it is safe and never get burned where it is not. The people who get hurt are not the ones who trust AI too little or too much. They are the ones who trust it the same amount everywhere, and never stop to ask what a given mistake would actually cost.
So for the AI-assisted work on your own plate: are you trusting it one task at a time, or have you quietly settled on a single level of trust for all of it?