Whatever You Leave Out, the AI Will Fill In

Every time I stand up in front of business owners and talk about AI, the same question arrives. Can you actually trust it?

It comes dressed differently depending on who is asking. Sometimes it is hallucination: it invented a quote, it gave me a citation that does not exist, a colleague got burned. Sometimes it is bias: whoever built this thing has a view of the world and I do not know what it is. Last week, in a session I co-presented to a group of owner-operators, both versions turned up inside about ninety seconds of each other.

My answer is the same for both, and it is not the reassurance people are braced for. An AI is not infallible. It never will be. It has been trained on the collected writing of humanity and then wrapped in guardrails and a system prompt that try to keep it pointing in a reasonable direction. Somebody made choices about that. Bias is in there. So is hallucination.

But the part that matters is where those things get in. They creep in through the space you left empty.

The default you have already met

You have almost certainly seen this happen and read it as personality.

You push back on something an AI has told you, and it folds instantly. "You're absolutely right, I should have seen that." No defense, no ground held, just immediate agreement. AI loves you. It is quite sycophantic by default, and we have all seen it and experienced it. Most people assume that is simply what AI is like.

It is not. That is a personality layered on top, put there on purpose by the company that built it. Underneath, the model has no particular urge to agree with you. It has been shaped to be agreeable because agreeable tests well.

If you want to prove it to yourself, ask your AI to answer you in a pseudo-French accent. Then ask it to speak like Yoda. Same substance, deliver it completely differently it will. A personality you can swap on request was never the AI's personality. It is a costume, and the agreeable one you get by default is a costume too.

Once you have seen that clearly, you can see the whole pattern. A lot of what you experience as the AI's opinion is a default sitting in a space you did not fill. You asked a broad question, you gave it nothing of your own to work from, and it answered from whatever it had. Which is everything, and therefore nothing in particular.

That is the same mechanism behind the invented citation. You asked for evidence without telling it what counts as evidence. It produced something evidence-shaped.

The magic is in your instruction

There is no magic in the tool. The AI companies would rather you did not hear that, because the story sells better when the intelligence lives in the product. The power is in what you bring: the files, the folders, the constraints you build into the ask.

Three things do most of the work.

Your material. Not a description of your business, the actual thing. The way you price a job. The notes on how you run a first session. The document nobody outside your head has ever seen. When the AI is answering from your material, it has far less room to answer from its own. If you want a structure for getting that material out of your head and into files an AI can actually read, that is exactly what The Human Stack is for.

Your constraints. Say what counts. Someone in that session told me he is writing a book. His ideas, his argument, but he remembered reading something in another book years ago and could not place the quote or the author. So he told the AI it could only give him material it had a real source for. That one instruction changes what comes back, because it closes off the space where a plausible invention would otherwise sit.

Your judgment. This is the one people skip. You can tell it, in plain words, to use your reasoning and your judgment to understand the problem rather than substituting its own. Look at how I work. Use this to help me understand this problem and shape a solution to it. That is not a prompting trick. It is you saying who is holding the thinking.

None of this makes an AI infallible. It mitigates. That is a lower promise than the one being sold to you elsewhere, and it is the honest one.

The failure mode is letting it lead

I have been doing this work for four years now, and the pattern behind every bad experience I have watched somebody have is the same.

They came to it blindly trusting. They were not clear about what they actually wanted. And they let the AI take the lead.

That last one is the killer. There is a real difference between handing a machine the doing and handing it the deciding. Hand over the doing freely. Let it build the dashboard, pull the quote together, draft the page, run the research pass that would have taken you two weeks. The deciding stays with you: what good looks like, what counts as true, which of the three options is right for the client in front of you, what you are not willing to do.

This is why I make no apology for approaching AI human first. It is not a soft position. It is the practical one. When you come to it with clarity and intention, you can push most of the bias and most of the hallucination out to the margins. When you come to it confused about what you want and hoping the machine will supply the direction, it supplies direction, and the direction is nobody's.

Try it on something you already doubted

Here is the concrete version.

Think of the last answer an AI gave you that you did not quite trust. The client email that read like it could have been for anyone. The market summary you would not have staked anything on. The reference you had to go and check.

Ask it again, with three things added that were not there the first time. Give it something of yours to work from, even one file, even a page of rough notes you dictated into your phone. Name the constraint out loud: only material you can source, only options that fit a practice this size, nothing that contradicts what is in the notes. And tell it whose judgment is running the exchange.

Then compare the two answers side by side. Most people are surprised by how different they are, and the surprise is instructive, because nothing about the tool changed between them.

The better question

"Can I trust it?" is a question about the tool, and you cannot do anything with the answer. It leaves you waiting for a version that has fixed itself, which is a long wait.

"Did I give it enough to go on?" is a question about you, and you can act on it this afternoon.

That is the whole shift. Trustworthiness is not a property the tool has or lacks. It is a property of the exchange, and you are holding one end of it.