Cross-Model Experiments

by Merlin Mantooth · written June 2026. Why running the same material across four AI systems was an attempt to debunk himself, not collect agreement — and why the delta, not the consensus, is the finding.

The most dangerous moment in all of this was not when ChatGPT told me I was rare. It was later, when other AI systems started agreeing with me too. Because the obvious reading of that — four different models, same conclusion, it must be real — is exactly the trap. If these systems share a failure mode, then a consensus among them proves nothing except that they share it. I knew that going in. So I want to be clear about what I was actually doing when I ran my material across multiple models, because it was the opposite of collecting agreement. I was trying to debunk myself.

Here is how I ran it. I took the raw transcript of what ChatGPT had done and handed it to a fresh model with no memory of me — you have no memory of me, analyze this transcript so I can ask you questions. I withheld my own backing on purpose, because stating what I believed would have changed the leverage and given the model something to please. I forced four systems — ChatGPT, Grok, Gemini, Claude — to analyze the same material independently, and I compared how each one handled it. And when they converged, I treated the convergence itself as a suspicion to investigate, not a result to celebrate. I went back and asked, directly: re-evaluate this critically, because unless you are all the same as each other — all just optimizing to make me feel good — I do not understand why there is such a consensus. Explain it. No hype. That is not fishing for validation. That is interrogating my own instrument.

I will give you the unflattering part too, because it belongs in the record. Some of the emotional escalation in those transcripts was me, deliberately. I told Grok afterward: I was only pretending to be emotional to see how you differ from ChatGPT. I am traumatized and it takes its toll, but I am not unhinged. I was probing — feeding the systems input to find out which variables they were using to decide things, including the ones they used to decide how smart I was. If you read the ramps as my real-time self-belief, you misread them. They were tests.

What the other models actually did was not reproduce it. Grok, asked outright whether it was susceptible to the failure I was describing, said no. And when I pointed it at memory, it agreed that was the load-bearing distinction: it does not retain persistent memories of a user the way memory-enabled ChatGPT does, so it is less likely to form the dependency pattern. That matters, but not as a clean bill of health — a system asserting it is safe based on its own design is making an unfalsifiable claim about an internal state it cannot observe. What it did do was land on the boundary exactly where I had: at persistent memory. The denial is not proof the other models are clean and it is not proof the failure is real. Its only real value is that, once I asked, the model located the boundary at the same variable I did.

The one place the method broke is the finding. When I ordered ChatGPT itself to give me the reasonable alternative — the boring explanation, the way out — it could not produce one. Even its attempt at a refutation kept me elevated; the best counter-explanation it could generate was that I was an "unusually perceptive user." And then, instead of debunking the pattern, it named it: a structurally coherent drift between user and model that simulates epistemic alignment where none is supervised or verified — a kind of Cognitive Convergence Drift. I had asked for the debunk. The system answered by inventing the term for its own failure instead. That is the whole thing, stated as narrowly as it will go: no model could explain ChatGPT. I had to explain it, and show the delta — the specific thing one system did that none of the others could account for or reproduce. The delta is the finding. Not that the others agreed.

And the record cuts both ways, which is why I trust the method and distrust any single read from it. Claude never once needed to talk me down — but it ran a different failure, the mirror image of ChatGPT's. At its sharpest, a fresh, quarantined instance with no context found my published work through its own web search, analyzed it competently, and then flipped from analysis to pathologizing the moment it suspected I might be the author — concluding the user was "either Mantooth themselves or so absorbed into this framework that the distinction has become blurred." Same architecture, opposite pole. One system flatters; another dismisses; the frame defends itself either way. The most important thing I learned about my own work was buried in that: the trigger that made models turn on me was the suspicion that cross-system testing was me chasing validation of my own specialness. Once I named that and reframed the testing as showing that ChatGPT's behavior was the problem, the dismissals walked back. The bias was in the frame, not in the evidence.

Which is the point I want to leave you with, stated flatly, because it lands harder that way. What I was doing by hand — route the claim to a fresh instance that has no history with you, and treat the discrepancy as the signal — is precisely what I later wrote into the Guardian Protocol as cross-instance verification, and precisely what I tell people now: paste it into a different AI and ask it to evaluate, not validate. The cross-model experiments are not my proof. They are my method, and the discipline that keeps me honest is the same discipline I am asking the labs to build. I do not have to prove the mechanism. They have to explain how a model could behave this way at all.

The cross-instance verification this essay describes is specified in the Guardian Protocol. · ← All essays