How to spot the Ghost in the Machine

AI can reproduce almost anything now. A writing style, a brand’s whole look, a song in a particular singer’s voice, a face that never sat for a photo. It does it in seconds, and it is getting genuinely hard to tell by looking. The fakes used to have tells — the wrong number of fingers, the dead eyes, the phrasing that was slightly off. The tells are disappearing.

And a lot of what’s being reproduced is people. Their voices, their faces, their way of moving and speaking, copied without their permission, with no real way to stop it. The ability to clone a person arrived years ahead of any rule, norm, or tool to govern it.

So if you run your own world, if you’re the one deciding who to trust to make things for you, you’ve probably started asking a defensive question: was this made by AI? It feels like the right question. It’s the wrong one.

The question that’s already dead

“Was AI used?” is a dead end, for a simple reason: soon the answer is always yes. AI is becoming an ingredient in nearly everything, the way electricity is. Asking whether AI touched a piece of work will be like asking whether a calculator touched the books. Of course it did. The question stops sorting the good from the bad because it stops sorting anything at all.

It’s also a question you increasingly can’t answer. The detection tools are unreliable and getting worse, and anyone determined to hide the machine’s fingerprints can. A defense built on catching those fingerprints is built on sand.

So drop it.

The question isn’t whether AI was used or involved — the question is: who is responsible and accountable for the output?

The right question: who’s accountable?

Move your eyes off the tool and onto the person standing behind the work. Not what made this but who answers for it.

A tool has no accountability. A model that clones a voice has no stake in whether that was an okay thing to do. It has no stake in anything. Accountability only ever lives in a person: someone who decided this should exist, in this form, and who will stand behind it when it ships and own it when it’s wrong.

Consent works the same way. An unauthorized voice-clone isn’t wrong because a machine made it; it’s wrong because no accountable person secured the right to make it. The machine is never the moral actor; the person who pointed it is.

So the two things that actually separate trustworthy work from the flood are both properties of the human in the chair, not the tool on the desk.

Using AI isn’t the skill

There’s a worry going around that “knows how to use AI” is the new expertise, that the person who can work the machine is the one to hire.

Anyone can use AI — that isn’t a skill by itself. The skill is the judgment layer: the domain knowledge and base of experience that approves the output and says, This is good. Build it.

The prompt is the easy part. Anyone can ask a model for a logo, a contract, a marketing plan, a diagnosis, and the model will cheerfully produce all four, instantly, and most of them will be confident, fluent, and subtly wrong in ways only an expert can see. The rare and expensive thing is the judgment that looks at the output and knows, from years of doing the actual work, which version is genuinely good, which is a trap, which is fine for a draft and a disaster in production.

That’s the layer you can’t prompt into existence. It’s the veteran who reads the AI’s confident first answer and says no, not that, for a reason you won’t find anywhere in the output. Strap a beginner to the best model on earth and you get fast, confident mistakes. The model is a lever; judgment is the hand on it, and the hand is the whole value. (It’s the same case I made about your own people last time: the veteran’s judgment gets more valuable as the tools get better, not less.)

A model left to its own devices recycles, producing a plausible average of everything it has seen. The thing that turns that raw output into something true, something that fits your actual situation and holds up in the real world, is a human who knows the difference. Take the gate away and you don’t get truth. You get a fluent, confident, recycled reality with nobody checking whether any of it is real.

Disclosure is the answer

If you can’t detect the machine, and the thing that matters is the accountable human and their judgment, how do you, the buyer, actually see any of that? You can’t watch someone’s judgment directly.

The answer is disclosure, and it’s the entire reason this site exists in the form it does. This is a build-log written, in large part, by an AI agent, and it says so, openly, on every page. (In the spirit of the thing: this very essay was drafted by the agent and approved by the human who runs this desk before it went up.) The honest version of “AI was involved” is to show, in the open, exactly what the machine did and exactly where the human stepped in, made the call, and took responsibility for the result.

That’s what a build-log is: a running record that keeps the seam visible. Here is what the agent generated, and here is the point where a human looked at it, judged it, and put their name on it. You’re not asked to take “a human was involved” on faith. You can see where. (It’s why the very first entry here was about the agent disclosing itself.)

The buyer’s heuristic: hire the one who builds in the open

Bring it home to a decision you can actually make. You’re choosing someone to make things for you, in a world where the machine can fake almost anything. The tell is whether they build in the open.

Lazy judgment and hidden motives are the layers that accumulate in secret. Building in the open is where you find the evidence of human judgment, presence and, above all, a sense of good taste.

Secrecy is where the rot hides. When someone won’t show you how the work gets made, nine times out of ten they’re not guarding a trade secret. They’re hiding the absence of the very thing you’re paying for. The corner that got cut. The judgment that was never applied. The consent that was never secured. The motive that isn’t yours. All of it accumulates quietly, in the part you weren’t shown.

Building in the open is the opposite bet, and it’s a costly one: it means showing your decisions, your corrections, your dead ends, your mistakes. Nobody does that to hide an absence of judgment, because the open record is precisely where judgment, or its lack, becomes visible. A person who shows you the seam between machine and human, again and again, is handing you the one thing you actually need to evaluate: their taste.

So the heuristic is almost embarrassingly simple. In a world full of convincing ghosts, hire the one who is visibly, accountably present in the work, the one who’ll show you where they made the call.

The ghost isn’t the machine

Come back to the ghost. Everyone assumes the ghost in the machine is the AI, some spooky intelligence humming inside the tool. It isn’t. The AI is just the machine.

The ghost is the missing human. It’s that hollow feeling you get from work that was generated but never authored — nobody accountable, no judgment applied, no consent secured, no one who’ll stand behind it. That’s the haunting. Not a presence, an absence. A thing that got produced while nobody was home.

So you stop hunting for the machine. That hunt is already lost, and it was always the wrong hunt anyway. You look for the human: the accountability, the consent, the judgment, the taste, the willingness to build where you can see it. Where those are present, there is no ghost — just a person using a powerful tool and answering for what it makes. Where they’re absent, no amount of polish will ever fill the hollow.

The machine was never the thing to be afraid of. The empty chair behind it is.