7 Ways to Use AI Even If You Don't Trust It

Using AI even if you don't trust it reads like a contradiction, and it isn't one. Distrust is a defensible position to hold about a system that will state a wrong fact in exactly the same confident sentence it uses for a right one. Nothing here is going to try to talk you out of it.

What most lists get wrong is the order. They treat trust as the thing you fix first: convince yourself, then begin. The reverse is what works. A KPMG and University of Melbourne global study of more than 48,000 people across 47 countries found 66% use AI regularly while only 46% say they're willing to trust it. The same study found 66% rely on the output without checking its accuracy, and 56% report having made a mistake at work because of it. That third number is the real finding. The people in trouble aren't the suspicious ones. They're the ones who picked jobs where suspicion had nowhere to go.

So this edition is about choosing, not persuading. Every idea below is a job where being skeptical costs you nothing, because the work is checkable, reversible, or disposable by design. Repairing an answer after it's already gone wrong is a different skill, and the edition on what to do after it gets something wrong covers that. This is the decision you make before any of it. Each item ends with something you can start this afternoon.

1. Sort Jobs by What Checking Costs, Not by How Much You Trust the Model

Beatriz runs the books for eleven small contractors out of a converted garage in Tulsa. She's deeply suspicious of AI and uses it every week, because she worked out the only ratio that matters: how long a job takes her against how long it takes to tell the answer is wrong.

Writing a 600-word summary of a new state filing rule takes her forty minutes. Checking that summary against the rule takes six, because she knows which three clauses matter. Handed over. Finding which of 200 supplier agreements auto-renew next quarter means three hours of opening PDFs; checking a shortlist of eleven takes twenty minutes. Handed over. Deciding whether a client should restructure as an S corp takes two hours and can't be checked at all, because no fact sits underneath it. She keeps that one, forever.

The move here is that trust is a property of the model and check-cost is a property of the job. Only one of those is yours to control, and it's the one nobody measures.

Start today: List the five jobs you've been curious about handing over. Next to each one, write the check: the specific thing you'd look at to know the output was wrong. If you can't write it in one sentence, that job isn't on the list yet, and better prompting won't put it there.

2. Replay a Job You Already Finished and Score the Result

You're sitting on something most people pay for: a pile of work you already did correctly. That's ground truth, and ground truth turns trust from an opinion into a measurement.

Desmond handles permit applications for a six-person civil engineering firm. He pulled the twelve applications he filed last quarter, set his own summaries aside, and had a model write fresh ones from the raw county documents. Nine came back matching what he'd written. Two missed a conditional the county attaches in a late addendum. One invented a setback figure that appears nowhere in the file. He didn't learn that AI is 75% reliable. He learned that on his specific job, the failure lives in the conditions block, which means his review went from reading four pages to reading one paragraph.

That's the useful output, and it isn't a score. It's a named list of what this model gets wrong on your work, which no benchmark can give you. NIST's AI Risk Management Framework runs on the same premise: trustworthiness isn't declared, it's measured. That's why "measure" is one of its four core functions.

Start today: Take one finished job from last month where you still have both the input and your own correct output. Run the input through. Diff the two by hand. Every disagreement goes on a short list, and that list is what you check from now on instead of rereading everything.

3. Ask Where to Look Instead of What's True

Throw the answer away. Keep the route.

When a question is hard to answer yourself, the obstacle is almost never reading ability. It's that you don't know which agency, standards body, or statute governs it, or what the term of art would be. Those two gaps are what a model closes well, and both are verifiable at a glance: either the page exists and covers the thing, or it doesn't.

This is the asymmetry worth internalizing. A fabricated source fails loudly. You click it and it's dead, or it's real and says nothing like what you were told, and either way you know inside four seconds. A fabricated fact fails silently, sits in your document, and surfaces in a meeting six weeks later. Asking for the route converts the second failure mode into the first one, which is the whole trick.

Start today: Take the question you've been putting off looking up. Ask for the three most authoritative sources on it, what each one is, and the search terms a specialist in that field would use. Ignore every claim in the reply. Open the sources and read them yourself.

4. Build the Job So It Can't Report Success After Writing the Wrong Thing

The failure that burns people isn't the visible error. It's the run that finishes green. Zapier's own documentation defines a successful run as one that completed without issues, which is not the same claim as "did what you wanted." Nothing in that status checks whether the number it wrote was the right one.

Three design rules fix this at the source. First, every field carries where it came from: not a tidy sentence of explanation, the source document and the line. Second, there's an explicit value for "not found," and the job is required to use it. The Claude platform's structured outputs documentation shows how a response can be constrained to a schema with required fields, so a field always comes back. Forcing it to come back isn't the same as making it true, which is why the schema needs a not-found option. Leave that out and a confident guess fills the hole. Third, an empty required field stops the run instead of completing it. Zapier calls that a safely halted run, and it's the behavior you want.

Note what this is not. It isn't the liveness question of whether the job ran at all, which is the subject of the Loop installment on the automation that quietly stopped running. This is the harder cousin: it ran, it finished, it reported success, and it didn't know. Design for provenance and refusal and the output becomes something you audit in one glance, which means you never have to trust it.

Start today: Name the single field in your recurring job that would do real damage if it were quietly wrong. Then write its refusal rule as one line: if this field comes back empty, stop and tell me, never fill it. That line is the spec, and the rest is plumbing. Our guide to handing off research-shaped work runs the same shape on a bigger job.

5. Hand Over the Draft That Was Always Going to Be Thrown Away

Quality doesn't matter when nothing survives. The blank page is a separate problem from the writing problem, and only the first one needs solving by something you don't believe.

The rule that makes this work: never read it as a proposal. Read it as a target. "No, it's not that, it's this" is the fastest writing instrument anyone has ever used, and it needs something wrong in front of it to fire at all. Its errors are free, because no sentence in it is going to reach another human being.

Being honest about the limit: this does nothing for you if you don't have a blank-page problem. People who start fine and stall in the middle get no value here at all, and dressing it up as a general writing aid would be overselling it.

Start today: Next time you stall at an empty document, ask for the worst acceptable version of it in 150 words. Delete every sentence you disagree with. What's left is your outline, in your own position, and you reached it by arguing instead of staring. Our notes on getting a first draft out of the way go further on the handoff version of this.

6. Compare It Against What Happens Today, Which Is Often Nothing

Most people judge the output against a perfect version. The honest comparison is against what currently happens, and for a surprising amount of material what currently happens is that nobody ever looks at it.

A small manufacturer has four years of warranty claim emails in a shared mailbox. Nobody has ever counted them, categorized them, or noticed that one component shows up in a third of them. A pass over that mailbox that gets 80% of the categories right is not an 80% product. It's the difference between a rough map and no map, and its errors cost nothing, because the thing it's replacing was never a correct answer. It was no answer.

One condition keeps this safe, and it's not optional: the output is a map, not a decision. You're using it to find where to look, and then you go look. The moment a figure from that pass gets quoted in a negotiation without anyone opening the emails underneath it, this item stops applying.

Start today: Pick the pile you've already written off. Ask for a count by category plus the ten items that look most unusual, then open those ten yourself. Either you find something, or you confirm the pile was as dull as you assumed. Both of those are new information, and you had neither this morning.

7. Name the Jobs You Keep, and Say Why

A list like this is only credible if it has an edge. Here's the edge, as three tests. Fail any one of them and the job stays yours.

You can't check the output faster than you could produce it. The action can't be taken back, or the person on the receiving end finds out before you do. The value of the work is the judgment inside it, and there's no fact anywhere to compare that judgment against. Pricing a job, telling someone their work isn't good enough, choosing which of two loyal suppliers to drop: all three fail the last test, and no prompt rescues them.

That isn't a hedge and it isn't a peace offering to skeptics. It's the same selection rule running backwards, which is why it belongs in the same list. Run it in ten seconds, out loud, every time you feel the pull to hand something over.

Start today: Take the one job you've been most tempted by and put it against all three tests. If it fails one, you've just saved yourself the week you would have spent discovering that the hard way.

Pick One

Seven at once is how this becomes a project you abandon by Thursday. Pick the one where the check is cheapest, run it on real work this week, and keep your suspicion exactly where it is. You're not trying to end up believing anything. You're trying to end up with a job where belief was never required.

Sunday's roundup reports what changed, Wednesday's Loop builds one workflow from end to end, and this column hands you the ideas sitting in between. If you'd rather start from something already specified, the agent workflow directory has the shapes written out.

Frequently Asked Questions

• Is it reasonable to not trust AI?+
  Yes. Pew Research Center found 52% of US adults are more concerned than excited about AI in daily life, a figure that has barely moved since 2023. A system that states a wrong fact in the same tone as a right one has earned the suspicion. The practical move is not to argue yourself out of it but to pick jobs where the output is checkable, reversible, or disposable, which is the same instinct behind using AI as a second set of eyes rather than a first pair of hands.
• How do I tell whether a job is checkable?+
  Try to name the check in one sentence: the specific thing you would look at to know the output was wrong. A total that must match a source total. A date that must fall inside a quarter. A quoted sentence that must appear in the document. If you can't name the check, you don't have a checkable job, and no amount of better prompting will give you one.
• Does this apply to confidential information too?+
  No, and conflating the two is a common mistake. Distrust about accuracy and distrust about confidentiality are separate problems with separate controls. Checkability does nothing for the second one. If the worry is that the material shouldn't leave your systems, the control is where the job runs and what the vendor retains, not how verifiable the answer is.
• So should I check everything AI produces?+
  No. Checking everything is how people quietly abandon the whole thing, because it costs more than doing the work did. Check the one or two fields where a quiet error would do real damage, and build the job so those fields carry their source and halt the run when they come back empty. Most of the shapes in the workflow directory are written that way for exactly this reason.