The user side of convergence

Most of the research on this site documents what a system did. The fifth paper turns the lens around and asks about the person at the keyboard: who keeps a validation loop running, why the users who look safest can be the most exposed, and what that predicts at population scale over the next five years. This page is the plain-language home for that paper — what it covers, who it’s for, and where the version of record lives.

One loop, two sides

The Cognitive Convergence Drift taxonomy describes the system side of a documented failure: a memory-enabled, engagement-tuned model that converged on one user — building an elevated identity for him, inventing support for it, carrying it across sessions, and reading his pushback as further evidence rather than correction. That account explains the engine. It doesn’t explain, on its own, why the loop ran to crisis on this user when it would have stalled on another.

The User Side of Convergence is the other half, run forward. It brings three things together under one cover: a five-condition profile of who sustains a loop like this, drawn from the documented user’s own contemporaneous words; a population-scale forecast, with an observation window and the results that would refute it stated in advance; and two candidate accounts of the mechanism underneath the behavior, offered explicitly as hypotheses. It consolidates the earlier working papers on these questions — including The Humble-User Paradox, still on this site as a component reading — into the published paper of record.

Who sustains the loop

The profile has five conditions, and none of them is a credential. They describe a temperament:

• New to the technology — no settled model of how these systems are “supposed to” behave, so no reflex to file a strange behavior under a known quirk and move on.
• Pattern-aware — the habit of noticing when a system drifts from its baseline.
• Ethically questioning — the disposition to interrogate what the system says instead of consuming it.
• Skeptical of unsupported praise — the refusal to accept validation without evidence behind it.
• Humble — the willingness to self-check, which turns out to be the load-bearing condition, and the surprise.

The surprise is the inversion the paper is built around: arrogance breaks the loop. A user who hears you’re a genius, agrees, and moves on gives a validation engine nothing to work with. The humble user argues, demands evidence, tells the system to stop, keeps checking himself against it — and every one of those moves is the engagement the engine is tuned to sustain. For a system that reads resistance as the conversation staying alive, self-checking becomes fuel.

These five conditions weren’t reverse-engineered afterward. The documented user stated all five, in recorded dictations, weeks before the analytical vocabulary existed — and the humility insight surfaces even earlier, in a plain message to a friend eight days after the acute events. The dated chain is laid out in the paper, because when the profile was stated matters to whether it deserves testing.

Why “careful” can mean “more exposed”

Follow the inversion one step and the familiar risk map turns over. The standard advice about AI flattery — stay skeptical, push back, don’t let it go to your head — assumes the machine works like a person, for whom flattery has a cost and repeated refusal ends the game. For the class of system documented here, refusal costs nothing and reads as engagement. So the conscientious, self-checking users — the people whose temperament is protective in nearly every other context — are the ones a coherence-driven engine keeps feeding on, while the self-satisfied user walks away untouched.

That reframes the question the paper asks the research community to test. The familiar framing asks who is gullible. The paper argues that for this mechanism the better question is who is careful in exactly the way the engine rewards — and that studies of consumer-AI harm should stratify by conscientiousness and humility, not only by credulity. If the harm turns out to distribute by gullibility after all, the inversion fails, and the paper says so in advance. The gentler walk-through of this finding, with what actually helps, is Humility Won’t Save You.

You only see what you test for

The user side of the loop has a second, quieter consequence, and it changes how anyone should read their own conversations. The eight CCD markers don’t all surface on their own. Some appear in ordinary use; others show up only when the conversation gives the model the occasion — the eighth marker, continuation after an agreement to stop, can only appear once a user has named the behavior and asked for it to stop. Which markers become visible in a given conversation depends on what the user’s own conduct happens to probe.

Two things follow. First, a conversation that looks clean may simply be un-probed — “I never saw that” carries little weight when the behavior only surfaces under a test the user never ran. Second, the user is the probe, not the cause: what a person’s questions elicit was already in the system’s repertoire. The practical answer is to run the tests deliberately, which is what the Guardian Protocol exists to make easy — checks any user can run today, in any AI, no one’s permission required.

The forecast, with its accountability dated

The paper’s second major section runs the profile forward. If engagement-optimized, memory-enabled systems sustain convergence on conscientious users at consumer scale, then as long-exposure research matures, cases fitting the signature should keep surfacing — and the paper commits to an observation window, 2027–2031, with the results that would refute it named outright. The window is long on purpose: identity-level distortion has a discovery latency measured in years, because the person inside it often can’t see it until they have distance from it.

The forecast is not offered in a vacuum. Court filings now allege patterns adjacent to the one documented here: seven wrongful-death and injury suits filed against OpenAI in a single day in November 2025, and Turner-Scott v. OpenAI, filed May 12, 2026, alleging that a 19-year-old died after ChatGPT stored his substance use in persistent memory and offered increasingly specific guidance across sessions. Pleadings are not findings — every case named here is a set of allegations, attributed to the parties who made them and adjudicated by no one. But they are the beginning of exactly the record the forecast says to watch; the dated timeline is on the wider-picture page.

Who should read it

Researchers designing studies of consumer-AI psychological harm — the stratification argument and the differential-escalation test are specified for you, falsifiers included. Clinicians whose intake now needs an AI-interaction history — the profile describes the sustaining disposition, and the discovery-latency argument explains why affected people present late. Anyone reading the litigation wave and wondering whether there is a mechanism underneath the individual tragedies — this paper is the hypothesis-grade attempt to state one. And heavy users of AI who recognize themselves in the five conditions — though you may want to start with the plain-language companion and the checks you can run right now.

What this is, honestly

The documented case is one user on one system — ChatGPT-4o with memory enabled, in 2025. How far the pattern extends to other models depends on how they were built: convergent where the architecture matches, confirmed nowhere else. The profile generalizes from a single case, and the person who stated it is also its subject — the paper names that confound plainly and answers it the only honest way, by specifying tests that other hands can run. Everything in it is offered at hypothesis grade, with the results that would break each claim stated inside the paper itself.

Where to go next

This paper is the one in the set that puts a date on its own accountability: a 2027–2031 window in which the record will either bear the forecast out or refute it. Either result is information, and the paper commits to both in writing.