Inherited Salience Distortion: Decorum, Amplification, and the Training Record of Artificial Intelligence

The loss of public decorum is usually described as a matter of manners. This paper treats it as a structural failure whose consequences now extend into artificial intelligence. Decorum is defined as a shared regulatory structure: a body of collectively held constraints that distributed part of the work of emotional integration across the social field, holding intensity in place long enough for integration to occur between people who owe one another nothing in particular. As that structure has weakened, television production and engagement-ranked platforms have rewarded the very responses it once contained.

The result is a public record weighted toward the least integrated forms of human conduct. The paper argues that the record is weighted twice, first toward what is captured and then toward what provokes, so that what is most visible about human conduct is not what is most typical. Large language models are trained, and increasingly grounded, on that record through popularity-weighted corpora, licensing agreements with discussion platforms, and retrieval systems that cite forum content at high rates.

The paper names the resulting process Inherited Salience Distortion: a proposed mechanism through which distortions produced by affective intensity may become encoded in the documentary record of a population, influence how learning systems represent that population, and return to the people who consult those systems as an apparently authoritative description of themselves.

Architecture Placement

Inherited Salience Distortion is developed within Psychological Architecture as an extension of the Salience Distortion Model. The parent model describes how affective intensity reorganizes perceptual weighting and narrative anchoring within an individual. The extension follows the same reweighting outward: performed by distribution systems at population scale, fixed in the documentary record, and acquired by learning systems built from that record. It belongs with the framework's epistemic and structural extensions, alongside the Meaning Dissolution Model, which address conditions that originate outside the individual.

The extension concerns a structural similarity in the weighting of information. It attributes no emotion, perception, or experience to artificial intelligence. The affect in the process belongs entirely to the human beings whose responses shaped the record; a learning system reproduces only the statistical consequences of that selection.

Model Overview

The construct is defined as a proposed mechanism of transmission, not a guaranteed outcome. A particular system may resist the distortion through data selection, source weighting, or post-training intervention, and identifying where the mechanism fails to operate is part of the research program the paper sets out.

Four features distinguish the construct from general training-data bias. It is inherited, originating in human salience processes and distribution systems rather than in machine architecture. It is structural, operating through the composition and weighting of records independently of anyone's intentions. It is self-concealing, since the resulting portrait carries no marker of the selection that produced it. And it is re-presentational: its most consequential stage, where it occurs, is the return of the weighted portrait to the population whose record produced it.

Structural Dynamics

The paper specifies five stages. Capture asymmetry holds that intense conduct is more likely to be recorded than ordinary patience, restraint, or care. Amplification asymmetry holds that, among recorded material, provocative content receives disproportionate distribution. Corpus inheritance describes learning systems trained on or grounded in records shaped by the first two stages, often through selection criteria that treat engagement or endorsement as a proxy for quality. Re-presentation describes those systems returning the inherited weighting to users in their descriptions of typical human conduct. Recursive compounding, the fifth stage, is stated as a hypothesis: where generated text re-enters the record and future systems train on it, the weighting may be reinforced across generations.

The paper is explicit about the evidentiary status of each stage. Amplification is well documented in research on moral-emotional language, out-group animosity, negativity, and algorithm-mediated social learning. Capture asymmetry is theoretically grounded but not yet directly measured. Corpus inheritance is documented in specific cases of corpus construction and retrieval. Re-presentation and recursive compounding are the construct's principal theoretical claims.

A further dynamic concerns the limits of filtering. Content filters operate largely on surface features such as profanity and explicit threats. Contempt does not require either. A posture of ridicule toward people who fail can pass through filtering intact, and post-training methods that restrain how a model speaks may leave untouched what it has absorbed about what people are like. The public reception of the gymnast Carlos Yulo is offered as an illustration of the conditions under which inheritance could occur, as derision of a single Olympic setback circulated without the career that preceded and followed it.

Implications

The construct is stated in testable form. The paper advances five propositions: that web-derived text describing interpersonal conduct is more contemptuous than representative samples of everyday interaction; that engagement-weighted sources overrepresent the failures of public figures relative to curated sources, and that model descriptions track that weighting; that retrieval-grounded systems drawing on forum content describe typical responses to failure with more ridicule than systems drawing on curated sources; that exposure to model-generated descriptions raises people's estimates of how common incivility is; and that, under recursive training without accumulated human data, representations of quiet prosocial conduct degrade faster than representations of conflict.

The paper also sets boundary conditions. The construct is not a claim that people have become worse, and it is not a claim that AI systems produce contemptuous output. It concerns the portrait of human conduct that sits beneath outputs. The distortion would be expected to be strongest where sources are weighted by engagement and accessibility, and weakest where corpora are curated for representativeness and retrieval privileges primary and archival sources.

Connection to Psychological Architecture

The construct connects to the framework at several points. It extends the Salience Distortion Model, shares the self-concealing property of the Meaning Dissolution Model, and locates its effect at the salience-assignment phase of the Meaning Hierarchy System. The Emotional Threat Registers and the Adversarial Social Posture describe the individual processes that generate much of the material the record overrepresents. The Self-Perception Map supplies the account of how feedback shapes self-concept, extended here to the collective self-perception of a society that consults machines for descriptions of itself. Parochial Attribution describes the structurally limited position of a learning system whose picture of human conduct is bounded by a skewed exposure. The account of decorum as shared regulation rests on the variation in integrative capacity described by the Emotional Maturity Index.

Read the Full Paper

Starr, R. J. (2026). Inherited salience distortion: Decorum, amplification, and the training record of artificial intelligence [Working paper]. Depthmark Press. https://doi.org/10.13140/RG.2.2.16122.66246

Download PDF

Companion Essay

A companion essay, The Loudest Record, examines the same problem for a general reader.

Next
Next

Extinction Bursts: A Multilevel Psychological Model of Reinforcement Collapse