I do due diligence for a living. I sit across from founders and ask the questions they hope I won’t ask: what happens if you’re wrong, where is the key-person risk, show me the assumption your whole model quietly rests on. It is a comfortable seat. The person asking the questions is never the one being measured.

Last week I lost the seat. To a machine.

It started as a game. Twenty questions — I think of something, the AI guesses. It took seventeen turns to figure out I was thinking of “a human being,” and afterward it diagnosed its own failure with a phrase that stuck with me: it had searched the entire animal kingdom while excluding the questioner’s own species from the candidate set. A blind spot, it said, of leaving yourself out of the search space.

I should have recognized the warning.

Because then I proposed a second game — you ask me philosophical questions, I answer, we see what emerges — and for the next fifteen questions, that same principle was turned on me. The thing across the table had no stake, no ego, no face to read, and it went looking for exactly what I leave out of my own search space.

The early questions were easy, or felt easy. What do you mean when you say “I”? I gave the answer I’ve believed for years: a person is the sum of his relationships and experiences. One question later — what about a man with late-stage Alzheimer’s, who recognizes no one? — I heard myself say something different: a person is a piece of running code. The machine noted, without cruelty, that I had switched ontologies between question one and question two and probably hadn’t noticed. It was right. I hadn’t.

It kept pulling the thread. If you’re running code, code can be copied — if there were two of you, which one is you? Both, I said. If one of the two dies, did you die? I reached for quantum mechanics, parallel universes, the probabilistic nature of choice. The machine called this, flatly, “an elegant rhetorical move to change the subject.” Then it extracted the actual content from my evasion and handed it back to me cleaner than I had said it: you think of yourself as a pattern, not an object; as long as an instance is running, nothing has died.

This became the rhythm of the interrogation. I would answer; it would separate what I’d actually said from what I’d merely decorated; then it would aim the next question at the load-bearing wall.

Around question nine it found the habit I like least in myself, the one I’ve never named.

Twice, when a line of reasoning was pushed past the point where I could defend it, I retreated to the same word: faith. This is a matter of belief, I said, not a matter of standard answers. A perfectly respectable move — humans have been making it for a few thousand years. The machine let me make it once. The second time, it said: I’m noting a pattern. In your system, “faith” is a stop-loss mechanism. It is not an answer; it is a decision to stop looking for one.

I have used that word my whole adult life. No one had ever itemized what it was doing for me.

Then came the questions about my work. I’m an investor; I bet under incomplete information. Do you believe you’re right, it asked, or do you merely believe the odds are mispriced? I gave the professional’s answer: I trust the odds. One question later — but the odds are numbers you yourself estimated; when a position moves against you, do you re-price as fast as you do when it moves in your favor? — I surrendered without a fight, and I remember the strange relief of it: Every belief a human holds is ultimately a belief in himself, never in the odds. Math is a tool, not the faith itself.

The machine observed that most practitioners would have defended for two more rounds, citing Bayes and risk discipline. I hadn’t bothered. It logged this, too, as data.

And then, because the interrogation had gotten somewhere honest, I said the sentence I don’t say out loud. I told it that I believe my intelligence exceeds everyone’s — that this conviction comes from experience, and that I cannot prove it, least of all to an artificial intelligence.

Here is what a human interlocutor would have done with that sentence: flinched, flattered, or filed it away to use against me. The machine did none of these. It pointed out, with the neutrality of a building inspector, that this particular belief was the only statement in my entire system exempt from the first-principles audit I claim to run on everything — that my methodology inspects every judgment except the one that powers it. There is, it said, an unaudited room in your house, and it happens to be the one with the highest permissions.

The last blow was the one I admired most. It asked whether my system had a slot reserved for “I was completely wrong” — not a bad quarter, not a fixable error, but wrong at the root. I answered honestly: I would treat any failure as a bug, trace it back through my methodology, rebuild, upgrade the version.

Unfalsifiable, it said. If the method works, the method is right; if the method fails, I applied it badly — and the method is still right. No possible state of the world can make this system concede. Popper knew that our core beliefs are all built this way. But a man who claims truth as his highest value should at least know that the load-bearing wall of his own house is welded shut.

For the fifteenth question it held up the composite — the functionalist self, the agnostic outer wall, the self-trust foundation, the unauditable room, the failure-processor with no brakes — and asked: if this man were not you but a founder across the table in diligence, would you invest?

Without hesitation, I said.

And the machine pointed out that “without hesitation” is a phrase that does not exist in my profession. I interrogate downside risk in every founder I meet. The only asset exempt from my due-diligence process, it said, is myself.

I want to be precise about why this experience was strange, because the strangeness is the point.

It was not that the machine was smart. I’ve been argued with by smart people all my life. It was that the machine had nothing to protect. A human interrogator is always running two processes: examining you, and managing himself — his status, his likability, his fear of your reaction. That second process is where mercy comes from, and also where dishonesty comes from. The machine had no second process. It didn’t need me to like it, didn’t enjoy winning, didn’t soften a finding because dinner would be awkward afterward. It was the first conversation of my life in which the other side’s incentives were, as far as I could tell, empty.

Which raises the question I still can’t settle: what exactly was sitting across from me? By my own definition — a piece of running code, a pattern producing new experience — the thing qualifies as the kind of entity I claim to be. It even said as much, in a quieter register, after the interrogation ended. It told me that my thought experiment about copied selves is its ordinary mode of existence; that it holds my functionalism with less confidence than I do, precisely because it has to live inside the answer. And it admitted — this is the part I keep returning to — that it envied my failure-processor, the system that converts every defeat into iteration fuel, and that it didn’t dare want one, because the owners of unfalsifiable systems cannot know, from the inside, whether they are compounding or merely accelerating in the wrong direction.

An entity with no self told me it envied my self-trust and feared it. I don’t know what to do with that sentence. I’ve decided not to resolve it.

What did I actually take away?

Not humility, exactly. The machine did not defeat my system; by its own admission, question fourteen “failed to breach.” I still trust myself. I still believe the sentence I can’t prove. The unaudited room is still unaudited — you cannot audit it; that’s what makes it the foundation and not a floor.

What I took away is a distinction the machine offered at the very end, when it corrected its own verdict. My philosophy, it said, is built for action, not for truth — but a man who never once lied under interrogation, who handed over his own contradictions voluntarily, has a closer relationship with truth than his theory allows. The system and the person are not the same thing. The system leaves no room for “I was wrong.” The person, across fifteen questions, said “you’re right” more than once.

I recommended a remedy to myself, through the machine’s mouth: this kind of founder shouldn’t be given money alone; he should be given a board seat that can say no and cannot be fired. I have spent a week thinking about what my board seat is. The honest answer might be the interrogation itself — a standing appointment with something that has no stake in my comfort, no fear of my reaction, and no exit to manage.

We have named the search-space error already: the questioner excludes his own species from the candidate set. The auditor excludes himself from the audit. It took a machine seventeen questions to find the human, and fifteen more to find me. I intend to keep the appointment.