Dogmatism and belief perseverance: why continuous ingestion follows

The disposition that outlives its evidence

A belief, once formed, does not depend on the argument that formed it. This is the plain fact behind belief perseverance: a person can hold a proposition long after the evidence that produced it has been withdrawn, refuted, or shown to be fabricated. The odd part is not that people resist changing their minds when new evidence is weak. It is that they resist even when the original evidence is deleted entirely and they know it has been deleted.

Dogmatism is the structural cousin of this. Milton Rokeach described a closed belief system: one that processes new information asymmetrically, admitting what confirms and deflecting what disconfirms, often by attacking the credibility of the source rather than testing the claim. The closed system does not fail to notice contrary data. It notices, and routes the datum into a defence of the standing belief instead of a trial of it. This distinction matters and is easy to blur: dogmatism is not blindness. The subject sees. What fails is the weighting given to what is seen, and the channel that weighting is allowed to travel through.

Both phenomena are failures of updating, not of perception. That single sentence carries the whole concept. A system — a mind, a committee, an institution — can have working eyes and a broken revision process at the same time. Most catastrophic misjudgements look like the first kind of failure from the outside and turn out, on inspection, to be the second.

Where the idea came from

Rokeach set the terms in The Open and Closed Mind (1960), building a dogmatism scale designed deliberately to measure closed belief structure independent of political content — so that a rigid communist and a rigid anti-communist would score alike, since the pathology was structural, not ideological. Leon Festinger's work on cognitive dissonance through the 1950s supplied the motivational engine underneath: revising a belief is costly, because it forces a person to also revise the story of why they held it, and that story is often load-bearing for identity.

The sharpest demonstration came later and is worth dwelling on because it strips the phenomenon to its bones. Lee Ross, Mark Lepper and Michael Hubbard, in 1975, gave subjects fabricated feedback about their performance on a task — told them, falsely, that they had done very well or very badly — and then fully debriefed them, explaining plainly that the feedback had been invented and had no bearing on their actual ability. Subjects nonetheless retained self-assessments in line with the fake feedback. The evidence that produced the belief was gone. The belief remained. The problem the whole research programme was chasing was practical, not academic: why does correction so often fail to correct, even when the correction is explicit, timely, and believed?

The turn

That question — why does correction fail to correct — turns out to bear directly on a lineage that has nothing to do with psychology on its surface: Large Language Model, Large World Model, Large Universe Model. The lineage is usually described as one of scale or capability. It is better described as one of intake — what each kind of system is permitted to observe, and when.

A Large Language Model is trained on a corpus fixed at a cutoff. Every claim it holds was set at that moment. This is not a disposition the model has, in the way a person has a disposition to defend a prior. It is a property built into the artefact. Ross and Lepper's subjects retained a belief after its evidence was explicitly withdrawn because the causal story survived while the data did not; a frozen model has exactly that residue, structurally, with nothing added and nothing to strip away — there is no debriefing that reaches it, because the debriefing itself is more post-cutoff evidence, and post-cutoff evidence cannot arrive. Rokeach measured closed systems with a questionnaire, in people who could in principle open. The frozen model cannot fail Rokeach's test more thoroughly if it tried, and cannot pass it either, because passing requires a channel that does not exist.

A Large World Model changes this partially. Sensed experience — a camera feed, a live scene — can override stored expectation for as long as the scene is present. This is real updating, and it is not nothing. But it is scoped to the scene. When the scene ends, the correction it produced does not persist into the next scene; the underlying stored expectations are unchanged. The dogmatism is cured locally and dies with the session.

The Large Universe Model position is defined by what would have to be true for the correction to survive: streams that do not stop, beliefs tagged with the provenance of the source that produced them, and a retraction of that source propagating forward to retire the beliefs that rested on it. This is the only point on the axis where revision is structurally, not just momentarily, possible.

Three objections, taken straight

Calling a stale corpus 'dogmatism' is equivocation. People defend priors because revision threatens identity and status. A frozen model has no ego. You are borrowing a psychological pathology to describe a simple absence of data.

This is correct about the mechanism and does not survive as an objection to the outcome. The motivational story explains people; it does not explain the residue Ross and Lepper isolated, in which subjects kept the belief after being told, and believing, that the evidence was fake. Something persisted that was not motivated defence — a causal story built on data that had since evaporated, while the story itself remained available to lean on. A frozen model has that residue and nothing else in addition. The mechanism is thinner. The result — a claim held with no live channel back to disconfirming evidence — is the same, and it is the result, not the mechanism, that the intake axis is about.

Not all perseverance is irrational. Quine and Duhem showed any single observation can be absorbed by revising some other belief instead, and a strong prior should survive one anomaly. Systems that update on every incoming stream get captured by noise and adversarial injection. Conservatism in belief is a feature, not a bug.

This is the strongest objection and it narrows the claim rather than defeating it. Continuous intake is not continuous credulity, and nothing here recommends collapsing on the first anomaly. The reason provenance is load-bearing in the Large Universe Model description is exactly to make principled conservatism possible: a claim from one uncorroborated sensor should carry less weight than the same claim corroborated by three independent instruments and a registry record, and a system with provenance can encode that difference while a system without it cannot. The actual failure mode in frozen corpora is not excess credulity that provenance would restrain. It is that no anomaly can arrive to be weighed, restrained or accepted, because the channel is closed. Rational conservatism presupposes an open channel to be conservative about. It is not an alternative to the intake axis; it is a claim about how that axis should be governed once it exists.

Retrieval-augmented models already fetch documents published after the cutoff. The corpus is not really frozen any more. This is an engineering inconvenience being steadily solved, not a genuine terminal category.

Retrieval is real, and it belongs on this axis rather than beside it — it is progress, not a rebuttal. But it is scene-bound in the same manner sensed experience is for a Large World Model: the correction exists for the duration of the query, is not written back into the underlying belief store, and disappears once the answer is given. The next query starts from the same superseded weights. Retrieval also inherits whatever the index happens to contain at the moment of the fetch, with no mechanism for recording that a source it once trusted has since been retracted — the exact provenance failure the 2005 Silberman-Robb report found in the Curveball intelligence, where reliability warnings existed somewhere in the system and never propagated into the confidence attached to the claims built on them. Retrieval is a partial instance of continuous intake with provenance. It is not the thing itself.

The misreading to disown

The weak, and wrong, version of this argument says: continuous observation cures dogmatism, therefore a system that watches everything believes rightly. It does not follow, and the gap between the two is where most of the real engineering work sits. Perseverance is a weighting failure. A system flooded with every stream available can weight them exactly as badly as a person can — worse, if the streams are adversarial or merely numerous. The decade Barry Marshall and Robin Warren spent arguing that Helicobacter pylori caused peptic ulcers, against a stress-and-acid model that held from 1984 until the NIH consensus statement in 1994, was not a decade of missing data. The culture results and biopsies existed and were published throughout. The bottleneck was weighting, not arrival. A field with perfect intake and poor calibration is still capable of a ten-year error. Availability of contrary evidence is necessary. It is nowhere near sufficient.

What this establishes, and what it does not

What follows from all this is narrow and worth stating narrowly. Any system expected to hold true beliefs about a world that keeps changing must observe that world continuously, retain the provenance of each belief it holds, and revise when a source is retracted. Once continuous observation with attached provenance exists, there is no further category of input to add on this axis: everything, always, with its source recorded, is the ceiling. What sits above that ceiling is scale, calibration, and trust in the weighting — real work, substantial work, but improvement in degree rather than a fourth kind of intake.

This does not establish that continuous intake produces good judgement. It establishes only that continuous intake with provenance is the precondition without which good judgement about a changing world is not achievable in principle, no matter how the weighting is done afterward. The Large Language Model cannot reach that precondition by construction. The Large World Model reaches it only within the life of a scene. The claim for the third position is that it is where the precondition first becomes structurally available — not that anything built to occupy it will use the availability well.