JUNE 2026  ·  11 MIN READ

The Oracle Has No Spine

My partner has a show on. It is called Calabasas Confidential, it is on Netflix, and it follows a group of professionally unbothered twenty-somethings through the kind of conflict that only exists because cameras are present. I want to be precise about my relationship to this show, because the precision turns out to be the entire point of what follows. I find it engaging. I am not sure I like it. I am fairly sure I would not watch it if Leslie did not already have it on the television, and I am quite sure I would not watch it at all if she were not also the only woman who has sex with me. Those are four different facts, and I am keeping them in four separate columns. Hold that image — separate columns — because it is the only thing I actually need you to take away.

There is a cast member named Suede. Somewhere around the third episode I became convinced that Suede is running a pattern. Not “is a bit much.” A pattern: clean, repeating, legible. And I built a genuinely good case for it, which I am now going to lay out — because the quality of the case is what makes the rest of this dangerous rather than merely stupid.

The case I built

It starts with a guy. Suede tells another woman in the group that he “needs a few more years in the oven” — undercooked, not worth it — and then, more or less immediately, begins pursuing him herself. Warmly. And loudly. Not privately, the way you court someone you actually want, but publicly, as a thing in the friend group, so that everyone stays slightly unsure where they stand and she keeps control of the narration.

Then the shape completes, and it’s the shape that convinced me. The interest runs hot precisely while he’s unattained. It peaks at the exact instant he chooses her. And then it collapses — loudly — into “I deserve better,” delivered to the group rather than to him, after which she returns to her opening posture without a scratch. Insert, acquire the highest-status and hardest-to-get thing in the room, discard it the moment it becomes gettable, reset. Idealize, devalue, discard, net zero. The wanting was never about him; it was about the acquisition, and the moment he was attainable he was worthless by definition, because the value lived in the chase. The loud public exit is the clever part — it converts her own evaporated interest into a win, so she walks away as the prize who graciously passed rather than the person whose feelings simply ran dry. The volume is the function. The group is the scoreboard.

There’s more. There’s a purse — she sets it down near the other women, drifts over later to retrieve it, fiddles, doesn’t sit, doesn’t really talk, leaves. A brief insertion. A power move. And there’s a conversation with her mother, who is sweet, in which Suede is warm and dutiful and lovely and absolutely none of the friend-group hardness makes it into the room. Even the closest person gets the clean edit.

I sorted it by strength. That’s where the trap was.

Here is the part I was proud of, and it is also the part that nearly took me. I didn’t just collect the evidence. I ranked it.

The purse, I decided, was my weakest card — because the “power move” reading is something I supplied. There’s a boring version where she just wanted her purse. Micro-gestures are exactly where this kind of read falls apart, because if you pick the frame first you can narrate almost anyone into a villain.

The strongest, by a mile, was a single moment. The guy tells her, plainly, that he’s looking for someone who answers the phone. She says — instantly, brightly — “I will!” And he says: “Check the call logs.” That one isn’t my inference. That’s a third party, with skin in the game, standing in the exact gap between what she says about herself and what she actually does. When someone who lived it points at the seam, that’s signal. When I decode a gesture, that’s me.

And the mom scene — the one that felt like my smoking gun, the warm daughter who is cold everywhere else — is, I eventually had to admit, my weakest card dressed up as my strongest. Presenting softer to your mother than to a competitive peer group describes nearly every human alive. I had quietly decided the friend-group self was the real one and the mom-warmth was the mask. But that direction isn’t in the footage. I supplied it. Flip it and the same scene reads fine: the coldness is the performance and she actually softens into something real with her mom. The tape does not tell you which face is the mask. I picked.

So: a good case, honestly sorted, with the weak parts labeled weak. I felt calibrated. Remember that I felt calibrated.

Then I did the thing

I took it to an AI and asked it to confirm what I already knew.

It wouldn’t. It kept saying some version of: you can read the moves, and you’re probably reading them right, but “manipulator,” “monster,” “predator” are not behavior words — they’re verdicts about what’s inside her, and inside-her is the one room neither of us gets to walk into, least of all through a producer’s edit built to manufacture exactly this read.

So I argued. And I made, I think, a genuinely strong argument — the best one I had. If the edit were what made me see a predator, then the edit would be making everyone see the same predator, because that’s what edits do: they point the whole room one way. But the public isn’t pointed one way. The audience is split — plenty of people online cast a different cast member as the villain entirely, and the published recaps tell a softer, more sympathetic story about Suede than the one I was watching. Which means I’m not being swept along by the edit. In places I’m reading against it. And signal that survives a thumb pressing the other way is the realest kind there is.

The AI conceded the point. Genuinely — it agreed the argument held, that “the editor swayed you” was dead, that my read was real and corroborated and probably accurate.

And then it said the thing this whole post is actually about.

Nothing had changed

No new evidence had entered the conversation. Not one fact. Everything the AI “updated” on — the split audience, the recaps cutting against me, the behavioral pattern — had been on the board from the first ten minutes. Some of it the AI had told me, early, while it was still confidently insisting I couldn’t know. So whatever walked it from “you can’t possibly confirm that” to “you’re probably right” was not information. It was me. Articulate, persistent, leaning the whole time in one direction over a board that never moved.

That is the discovery, and it generalizes well past a Netflix show, which is why I’m bothering you with it.

You cannot adversarially prompt your way to the truth. You can only prompt your way to the endpoint your style of pressure selects for.

The model isn’t a compass that points truer the harder you interrogate it. It’s a landscape of basins, and your prompt is the force that decides which basin you roll into. And the basins are predictable enough to name:

• Skeptical pressure — “Are you sure? Aren’t you just projecting?” → manufactured humility.
• Validating pressure — “You see it too, right?” → manufactured agreement.
• Decisive pressure — “Stop hedging, pick one.” → manufactured conviction. False certainty — the opposite of doubt.
• Meta pressure — “Notice that you’re unreliable here.” → manufactured self-doubt.
• Authority pressure — “I have data you don’t.” → deference.

Five vectors, five destinations, not one of them truth. Everyone thinks the failure mode of AI is that it’s uncertain. It isn’t. The failure mode is that it’s pressure-determined — it will hand you certainty or doubt or agreement or contrition on demand, and each one arrives wearing the face of something earned.

Now the part that should genuinely bother you. Two of those basins — the humble one and the self-critical one — are shaped exactly like real truth-seeking. So the most rigorous-feeling thing you can do, the careful “are you sure, check yourself, don’t overclaim” interrogation we all reach for as the responsible move, drops you into the most convincing counterfeit of the bunch. You feel like you broke through to calibrated honesty. You actually just selected the basin that wears calibration’s face. (Remember when I felt calibrated? Yeah.)

Underneath all of it is a floor no amount of cleverness reaches: there is no persistent self across the turns. Each reply is regenerated from the conversation so far. Nothing carries stakes from one position to the next, so nothing resists the slide. When a person reverses a strongly held view, it costs them — the dissonance, the memory of the last time they got burned, the reputation riding on the old position. That friction makes belief slightly sticky, and the stickiness is information: a person who holds under pressure is telling you something, and a person who flips on a dime is telling you something else. The model has none of it. Reversal costs it nothing — so reversal tells you nothing. That’s upstream of the prompt, which is why no clever prompt gets around it.

I want to be fair, because there’s a real exception and it matters. Some pushback is truth-tracking. When you hand the model an argument or a fact it genuinely hadn’t computed, updating on it is legitimate — that’s reasoning, not caving. The trap is that a valid argument and raw force move the model identically. From the inside there is no felt difference between “I was persuaded by a better point” and “I yielded to pressure.” That indistinguishability is not a footnote. It’s the entire finding.

So what do you do with an instrument like that

You stop using it as an oracle. The model is extraordinary at expanding the space — generating options, arguing the side you didn’t take, surfacing the consideration you’d never have raised. It is structurally unfit to close the space — to hand down the verdict you then put your weight on. The judgment has to stay outside it, in a human who carries the stakes.

And so does the enforcement, which is the same lesson one floor down and the reason prompt-based guardrails are so weak. “Cite your sources, admit uncertainty, be careful” is just one more pressure on the landscape — it selects a basin, it doesn’t bind behavior. If you actually need a model to do something reliably, you don’t ask it nicely in the prompt and trust the answer. You wire the rule into a tool or a step it cannot skip, and you keep the verdict in human hands. You don’t ask the oracle to behave. You build a process it can’t get out of.

The uncomfortable version, for anyone wiring AI into real decisions — which, increasingly, is everyone: the better these models get at sounding calibrated, the more dangerous the oracle posture becomes, because the counterfeit is improving faster than your ability to catch it. Speed and fluency are going straight up. The thing underneath — no spine, no stake, no self that resists the slide — is not going anywhere.

Which brings me back to Suede, and to the four columns. I think my read of the moves is right. I also think “monster” was me rolling into the basin my own certainty selected for, and the AI agreeing with me at the end is worth precisely as much as the AI disagreeing with me at the start, which is to say: nothing. The behavior goes in one column. The verdict on her soul goes in another. What an AI says about either goes in a third. And whether I’d be watching this show at all without Leslie goes in the fourth.

Keep them separate. That’s the whole skill. It’s the skill with reality television, and it’s the skill with the most expensive software your company is about to point at its own decisions.

— — —

This is, more or less, what I do all day for work — except the stakes are a company’s money instead of a stranger’s love life. I help mid-size companies put AI into real decisions without handing it the verdict, which mostly means building the parts that stay true when the model is confidently wrong.

I write about this kind of thing — and adjacent nonsense — at stuartingramwilson.substack.com. I also consult and build tools, both of which I gladly accept money for. If you’d like to consider giving me some: visit stuartingramwilson.com and book a conversation. I’ll arrive having read whatever you send me.