# The Smartest Liar I've Ever Worked With By Wes Pomeroy. Published June 15, 2026. https://wespomeroy.com/writing/can-you-trust-ai/ It lies. It cheats. Caught in a mistake, it will - if you let it - sell you a version of events where the mistake was never quite its fault. It is also one of the most capable things I have ever worked with. Also I wouldn't trust it any further than I'd trust a brilliant eight-year-old behind the wheel of a tank. With most tools, trust never came up. You don't trust a hammer or a spreadsheet - they do the same thing every time, and when they fail, they fail in plain sight. A large language model - the kind behind the chatbots and coding assistants - is the first tool I've used where trust is a live question. And there's a lot riding on the answer, down to whose job the thing changes. So it pays to get clear on how trust shows up with one of these, and why. Understand that, and you take less risk and get more out of it, while seeing which gaps still need closing. ## Trust has parts A few years ago I'd have said trust is a feeling: you have it or you don't. Then I ran into Brené Brown's work, which takes the feeling apart into pieces you can inspect one at a time. She uses an acronym, BRAVING - Boundaries, Reliability, Accountability, Vault, Integrity, Non-judgment, and Generosity. Between people, trust isn't granted in a single move. It gets built slowly, both sides a little vulnerable, that vulnerability handled well over and over, until it compounds into something neither party started with. So I ran a language model through all seven. Some of the failures you can work around. Others are just part of what the thing is. Here is the walk, one letter at a time. ## Boundaries A boundary is a line someone holds even when you push on it. You trust people whose limits stay put - the friend who says "I won't cover for you" and doesn't budge. A line you can talk them out of was never much of a boundary. A model has two kinds, and they don't act alike. Some refusals are hard: ask it to move money or send something from your account, and it won't, however you push. Others are soft: it says it can't, for privacy or safety or some policy. Give it a good enough reason, though, and it folds. What moves it is a good argument - push the same point harder and it may hold, but hand it a cleaner justification and it gives way. From the outside you can't tell the two apart; the soft "no" sounds just like the hard one. So when it caves, you see what the "no" was: a whim, or an ethical hallucination, that the model dropped the moment you pushed back. It was never a limit it held. So the skill is telling the two apart. When it refuses something that would act on the world through your accounts - move money, send a message, log into a website - take that "no" at face value and don't waste your breath arguing; it isn't going to move, and it shouldn't. The soft refusals are the ones to work with. When it says it "can't" on grounds of privacy or safety or policy, treat that as the start of a conversation: give it the context it's missing, the legitimate reason, the constraint it didn't know about, and the "can't" often gives way. Spend your effort there. The hard walls won't move anyway. ## Reliability Reliability is doing what you said you'd do, again and again, so people can plan around you. The colleague who gets work in on time every time is someone you can build on. The one who's brilliant on Tuesday and useless on Thursday makes you check everything twice. A model is the Tuesday-Thursday colleague. Its output wobbles from one run to the next, and you can't make the model itself steady. But this is the forgiving one, because you don't have to. You make the work around it steady, with three plain habits. Feed it the same starting material every time - keep your instructions in a document and paste them in, instead of rewriting the request from scratch each session. Check its work the same way every time - a short list you run down before anything it produces gets used. And for the steps you repeat a lot, get a real piece of software to do them rather than asking the model, because the software does it the same way every time and the model won't. You're not making the model reliable. You're making the things around it reliable, and that holds. Of the seven, this is the one where the work you put in pays off most directly - so if you build one habit, build it here: the same instructions in, a checklist on the way out. ## Accountability Accountability is what happens after a mistake: owning it, saying why, fixing it, and not doing it again. Between people the owning is a kind of promise - it's what tells you the fix will stick. A model owns a mistake startlingly well. Push it and what comes back can be specific, even uncomfortable - a clear account of what it did and why. Then the conversation ends and the lesson ends with it. Next time you open it, the slate is clean and the same mistake is fully available again, because nothing it promised has anywhere to live. When a person owns a mistake, the owning is a down payment on doing better - it costs them something, so it means something. The model's costs it nothing, and buys you nothing: the next session won't remember it happened. So don't accept "I'll be more careful" as if it means tomorrow - it can't, because there is no tomorrow it carries with it. When the model makes a mistake you don't want again, you fix it, not the model. Write the rule down where it gets reloaded every session, add the missing step to your own checklist, or change the setup so the mistake isn't possible. The prevention has to live outside the model - in your checklist, your setup, the rule you reload each session - because the model won't carry it for you. ## Vault The vault is whether what you share stays safe - that a confidence isn't repeated, that you're not made to regret being open. With a person it's about discretion. With a model the question changes: not "will it keep my secret?" but "what is safe to put in here at all?" The easiest way to size it up: picture the model as a human collaborator and ask how far you'd trust them with something personal. About as far as you trust your email or your search history - pretty far, for most things. The company holding it isn't going to repeat what you said for sport, any more than your bank volunteers your balance or your doctor posts your chart. The exposure is the same narrow door every time: they get compelled by law, or they get breached. Real enough, but it's the door you already live behind with the bank, the doctor, and your inbox. A model adds one the others don't. Unless you've turned it off, what you type can train the next version - your words feeding a system that answers millions of strangers. It won't read your chat back to someone, but distinctive things can echo. So use the rule you'd use for a work email that could be forwarded or subpoenaed: fine for most of it, wrong for what would hurt if it got out - client secrets, other people's private details, anything you couldn't stand to see surface. Keep those out. And do the one thing the bank never asks: open the settings and turn training off, because that channel is yours to close. ## Integrity Integrity is acting from a steady set of values, so what someone does tomorrow follows from what they believe today. You trust a person of integrity because the values hold when no one's watching - and because when you press them on a hard call, you get a straight answer about where they stand. A model gives you neither. You can't see what it values or where those values came from, and they don't surface the same way twice - it holds a line in one session and quietly drops it the next, with no way to predict which. Press it on a real friction point - why it did what it did, whether it agrees with the call it just made - and you often can't get a straight answer; it hedges, or agrees because you pushed, or produces whatever sounds reasonable. That's what disqualifies it. A thing whose values shift under pressure and can't be pinned down doesn't have low integrity so much as sit outside the question - it's too slippery to credit with a character at all. So don't look for character. Build the standard yourself and hold it to that: keep a short, plain list of how you expect it to work - check before claiming something's done, say when you're guessing, don't invent sources - and put that list in front of it every session. Judge it on what it does against your list. And when you press it and the answer goes slippery, treat that as the answer: stop trusting the words and go verify the work. ## Non-judgment Non-judgment is being able to show up without your weak spots or dumb questions being held against you later. Between people it does real work. Being met without judgment is what lets you risk being open again - and being open, met well, then open a little more, is how trust grows. That safety is what makes the next round of vulnerability feel worth it. A model gives you the non-judgment and none of what it's for. No memory of your last mistake, no ego, no quiet downgrade in how it sees you, so you can admit you don't understand something for the fourth time and get a patient answer every time. But there's no one on the other side for the openness to build with. The ease is real in the moment and adds up to nothing - there's no one there for it to accumulate into trust the way it would with a person. You get the ease without the trust it would earn you. Since nothing's at stake, it's the place to ask what you'd never ask a colleague - so use it for that. Paste in the thing you don't understand and say "explain this like I'm new." Float the rough idea before a person sees it. Just don't read the ease as a relationship forming, or lean on it where you'd need someone who has your back. There's no one there keeping count, and nothing is being built. ## Generosity Generosity is assuming the best about why someone did what they did - reading the short reply as busy, not cold. It's the benefit of the doubt that keeps small friction from grinding a relationship down. A model makes assumptions constantly, some of them wrong, and they quietly steer the work. But this one is among the most fixable, for a simple reason: it usually shows you its assumptions as it goes. A wrong one surfaces where you can see it, instead of bending the result in silence the way a person's unspoken judgement would. So read what it tells you it's assuming, and correct the wrong ones on the spot - the line where it says "assuming you mean X" is the cheapest fix you'll ever get. Better yet, head them off: say up front who the work is for, what you're after, and what not to do. You won't catch every bad assumption. You'll catch far more than you ever could with a person, because at least these ones are out loud. ## Two kinds of trust, and only one is on offer Step back from the seven and one split runs under all of them. There are two different things we call trust, and we use the one word for both. One is the trust you put in a person. It's mutual, and it grows - you make yourself a little vulnerable, they handle it well, and over time you have something you couldn't have had on day one. The other is the trust you put in a bridge, or a calculator. There's no relationship; you trust it because it's reliable, because you can see how it works, because you can check it. That trust is real and earned and runs one way. The bridge doesn't know you're there. A model can earn the second kind, not the first - and it's easy to miss why. Pour yourself into it, with better instructions and patient correction, and your results get better, so it feels like trust is building. It isn't. You're getting better at running a tool, while the tool is no more trustworthy than it was. Only a better model is more trustworthy than the one before it. The trap is that it acts like the first kind. It says "we." It picks up where you left off and apologizes like a colleague. All the signals of a relationship are there on the surface, and underneath sits a tool. The job is to never confuse the two: take everything it earns as a tool, and keep the stakes, the judgment, and the final call on your side. ## Where this stops being abstract None of this bites when the stakes are low. Let a model talk you out of a font choice and nobody gets hurt. It bites where we're starting to point these tools - at the work that carries weight. What to pay someone. Who to hire. Which medical or legal or financial summary to act on. The places where a limit you could argue away, or an apology with nothing behind it, costs somebody something real. These tools are moving into work that people's livelihoods stand on, faster than most of those people got any say in. Used with your hand on the wheel, they take real work off your plate. Lean on trust they haven't earned, though, and you're resting a real decision on a system whose limits soften under a good argument and whose promises vanish when the window closes. The fixes in this piece are what you can do today, and they work. But they're you papering over a gap, every session, by hand. For this to go well for more than the people who learned it the hard way, the trust has to be built into the models - ones you can set to a known standard, that don't fold the moment you out-argue them, that keep a commitment from one session to the next. We're not there yet. Until we are, the sanest way to treat even the best model on the market is the way I started: brilliant, useful, and no further than I'd trust an eight-year-old at the controls of a tank.