Argument in Brief
American public life has undergone a visible loss of decorum. Elected officials speak in profanity from podiums and official accounts; strangers fight over lane changes and the pace of a checkout line; the struggles of private people and the losses of athletes and teams become occasions for public mockery. This essay reads that loss as a structural failure. Decorum, whatever its historical abuses, performed a regulatory function: it held emotional intensity in check between people who owe one another nothing in particular, and its erosion leaves contempt with less to contain it. Television production and algorithmic distribution did not invent contempt, but they converted it first into a product and then into a weighting rule, assigning the greatest visibility to the most intense and least integrated forms of human response. Visibility, however, is not prevalence, and the gap between the two is where artificial intelligence enters. Language models learn from the written record of a society, and that record has been shaped by systems that reward the loudest material over the most considered. When a forum thread outweighs an academic record, and the mockery of a champion's single fall travels without the account of his career, the training corpus risks inheriting a distorted portrait of human beings. The question the essay pursues is what the next generation of systems will take human beings to be when the record it learns from is weighted toward their worst moments.
A legislator stands at a microphone and uses a word that, a generation earlier, would have ended a career and now produces a fundraising email. A driver signals, changes lanes, and is followed for three miles by another driver who intends to settle the matter at the next light. A customer in a grocery line, delayed perhaps forty seconds by a cashier still learning the register, records the cashier on a phone and posts the recording with a caption designed to humiliate. A person who has lost a job, a marriage, or a measure of health discloses the fact publicly and receives, among the condolences, a stream of ridicule from people who have never met them. A team loses a final, and its youngest player, who missed the decisive attempt, becomes a template for jokes that circulate for years.
None of these scenes is extraordinary any longer, and that is the subject of this essay. Each represents the failure of something that once operated quietly in public life: decorum, the set of shared constraints governing how people conduct themselves before others to whom they owe nothing in particular. The loss of decorum is usually described as a matter of manners, a coarsening of taste, a shift in what is considered acceptable to say. That description is accurate as far as it goes, but it misses the structure beneath it. Decorum was never primarily about manners. It was a regulatory structure, and its failure has psychological consequences that extend well beyond the people who shout at cashiers.
Those consequences also extend into a domain that did not exist when decorum began to fail. Artificial intelligence systems are trained on the written record of the societies that produce them. That record is not a neutral sample of human life. It has been filtered, ranked, and amplified by systems designed to capture attention, and those systems reward the very forms of expression that decorum once contained. The result is a question without precedent: what will a machine conclude about human beings when the record it learns from was selected, at nearly every stage, for volume over worth?
Decorum as a Load-Bearing Structure
Within Psychological Architecture, human experience is organized across four interdependent domains, Mind, Emotion, Identity, and Meaning, and psychological health is understood as a matter of coherence: the capacity of these domains to remain integrated when conditions place them under strain. Most of the structures that sustain coherence are internal. A person regulates affect, maintains a stable self-concept, and holds an ordering of what matters. Some of the structures that sustain coherence, however, are shared. They exist between people rather than within them, and they perform regulatory work that individuals would otherwise have to perform alone.
Decorum is one of these shared structures. It is defined here as the body of collectively held constraints on public conduct that limits how much internal state a person may discharge onto others who have not agreed to receive it. A driver who is cut off experiences a surge of threat and indignation; decorum is the external expectation that the surge will be contained, that the indignation will remain a private event rather than becoming a pursuit. A spectator who watches a rival team lose experiences pleasure; decorum is the expectation that the pleasure will not be converted into the public humiliation of a young athlete. In neither case does decorum eliminate the emotion. It holds the emotion in place long enough for integration to occur, so that the response reaching other people is the integrated response rather than the raw one.
The history of decorum is not innocent, and the argument requires that this be stated plainly. Its conventions have been used to preserve hierarchies, to dismiss legitimate dissent as incivility, and to demand deference from people with less power than those they were expected to defer to. A person could satisfy every convention of an earlier era while practicing cruelty within its permissions. The concern here is the regulatory function of restraint, which is separable from the particular conventions through which restraint was historically enforced and from the inequalities those conventions often protected. What has eroded is the function, and the function is what carried weight.
That is why decorum is properly described as load-bearing. The Emotional Maturity Index defines maturity structurally, as the degree to which emotional processes remain integrated under strain, including impulse modulation, tolerance for ambiguity, and the capacity to hold competing perspectives at once. No society can assume that every member possesses these capacities at a high level at every moment. Decorum compensates. It distributes part of the work of integration across the social field, so that a person having a difficult day, or a person whose regulatory capacities are limited, is still held within a structure that expects restraint and makes restraint the default.
When a load-bearing structure is removed, the weight does not disappear. It transfers. The erosion of decorum returns the full burden of regulation to individuals at the moment the surrounding environment has become more provocative, more crowded, and more saturated with stimuli designed to produce exactly the affective surges that decorum once contained. What follows is not a society of worse people. It is a society in which the same people, under greater pressure, are less held.
The Failure in Ordinary Places
The loss of decorum is most visible at the top of public life and most consequential at its base, in the ordinary encounters where strangers meet without any particular reason to regard one another with care.
The Public Office
Elected office once carried an expectation of elevated conduct that was partly ceremonial and partly structural. Across the political spectrum, that expectation has weakened. Profanity now appears in floor remarks, campaign rallies, televised interviews, and official social media accounts, and it is used less as a lapse than as a technique: vulgarity signals authenticity, combativeness, and contempt for an opposing faction, and those signals are rewarded with attention and donations. The structural significance lies in the modeling function of public office. Officeholders are among the most visible demonstrations of acceptable adult conduct. When they discharge contempt publicly and are rewarded for it, the shared structure loses one of its principal anchors. Restraint becomes optional, and an optional constraint carries no load.
The Road and the Register
On the road and at the checkout counter, the failure takes a different form. The Adversarial Social Posture describes a configuration in which other people are experienced not as neutral cohabitants of a shared space but as obstacles whose presence registers as interference. A driver organized by this posture does not experience a lane change as a lane change; the other driver becomes an adversary who has taken something. A customer organized by the same posture does not experience a slow cashier as a person learning a job; the cashier becomes a barrier between the customer and the next obligation.
The Emotional Threat Registers explain why such encounters escalate so quickly. As emotional intensity rises, interpretive range narrows and complexity tolerance falls. At low intensity, a driver can hold several readings of an event at once: the other driver did not see, was distracted, was hurrying toward an emergency. At high intensity, only one reading survives, and it is the reading of deliberate offense. A forty-second delay is a trivial cost, but once intensity has narrowed interpretation to a single frame, it registers as disrespect, and disrespect is the kind of event for which shouting, filming, and occasionally violence feel proportionate. Decorum once interrupted this sequence at the precise moment when the internal capacity for checking had been narrowed by intensity. With that expectation weakened, the narrowed interpretation proceeds more readily to action.
The Spectacle of Struggle
The most revealing failure is the mockery of people in difficulty. Mockery serves several functions: it bonds groups, punishes perceived transgression, entertains, and expresses rivalry. One of its functions bears directly on the present argument, and it is most visible when the target is someone already destabilized, a person who has disclosed an illness, a bankruptcy, a divorce, or a public failure, or an athlete or team that has just lost.
The Self-Perception Map describes self-concept as constructed from layered inputs, including relational feedback and comparison. Comparison with a person in visible difficulty offers an inexpensive form of self-elevation. The spectator who mocks the athlete who fell has not trained, competed, or risked anything, yet in the moment of mockery the spectator occupies the position of the one who did not fail. The ridicule purchases a temporary sense of standing, and its cost is transferred to the person being ridiculed. Mockery of struggle, in this function, is the conversion of another person's instability into one's own provisional coherence.
Decorum once made this transaction expensive. Public ridicule of the suffering carried social cost; it marked the mocker rather than the mocked. When that cost falls away, the incentive to perform the transaction rises, and the conditions that once discouraged it no longer apply.
Humiliation as a Product
Television production did not create the appetite for humiliation, but it industrialized its satisfaction. Reality television, in many of its most successful formats, is organized around the manufacture and display of moments in which people lose control, lose standing, or lose face. Participants are cast for volatility. Environments are engineered to produce conflict through scarcity, disrupted sleep, alcohol, and confinement. Editors then assemble the footage to maximize the visibility of the worst moments, compressing hours of ordinary conduct into seconds and extending a single outburst through repetition, slowed playback, reaction shots, and musical cues that instruct the audience on how to feel.
The Salience Distortion Model describes how affective intensity reorganizes perceptual weighting and narrative anchoring, altering what feels true under pressure. Editing performs this distortion deliberately and from outside. It assigns maximal salience to the moment of humiliation and minimal salience to everything that surrounds and explains it. A viewer who consumes enough of this material learns, without ever being told, that conflict is the natural state of human contact and that the fitting response to another person's humiliation is amusement. Participants learn a parallel lesson. The person who throws the drink is invited back; the person who apologizes is edited out. Production does not merely display contempt. It selects for contempt, and it demonstrates to a very large audience that selection for contempt is how visibility is won.
Amplification as Salience Distortion at Scale
Algorithmic distribution extended the logic of the edit to the entire public sphere and removed the editor. Platforms that rank content by engagement do not set out to promote contempt. They set out to promote whatever holds attention, and attention is held most reliably by material that produces strong affective response. Research on social media diffusion has found that moral and emotional language increases the spread of political messages, that posts expressing animosity toward an opposing group draw more engagement than posts expressing support for one's own, and that false stories travel farther and faster than true ones, carried in part by the surprise and disgust they provoke.
The result is an asymmetry. Considered speech, measured disagreement, careful correction, and ordinary kindness are not suppressed; they are outweighed. They generate less response, so they receive less distribution, so they generate still less response. Contempt, ridicule, and outrage generate more response, so they receive more distribution, so they generate still more. In an individual, the Salience Distortion Model describes affect reorganizing what is attended to and what feels true. In an engagement-ranked environment, the same reorganization is performed by an external system on behalf of millions of people simultaneously. This is salience distortion operating at the scale of a culture, and the Meaning Hierarchy System identifies what is at stake in it: salience assignment is the first phase in the construction of meaning, and what is assigned salience is what enters the hierarchy of what matters. When salience is assigned by an engagement metric, the moments a society would least choose to be defined by become the moments that occupy its attention.
Visibility and Prevalence
The distinction on which the rest of the argument depends is the distinction between visibility and prevalence. A person who encounters a hundred instances of cruelty online for every five instances of kindness has learned something true about an information environment. Nothing in that ratio establishes the ratio of cruelty to kindness in ordinary life. Most of ordinary life is never recorded. A stranger who waits without complaint, a driver who lets another merge, a person who swallows an angry retort, a neighbor who helps without making a performance of it: these events are common, psychologically significant, and almost entirely without durable trace. A single recorded outburst, meanwhile, can circulate millions of times.
The record, in other words, is weighted twice. It is weighted first toward what is captured, and what is captured is weighted again toward what provokes. A person scrolling through filmed confrontations and viral humiliations does not experience the feed as a selection of extremes from an environment that rewards extremes. It is experienced as a window onto what people are like. That confusion of the visible with the typical is a psychological error in an individual. In a system that learns from the record, it becomes a structural one.
The Record and Its Weight
Artificial intelligence enters the analysis here, because language models are built in substantial part from the record that amplification produced. Three processes are involved, and they must be held apart: what becomes visible online, what is selected and weighted during training, and what a deployed system subsequently says about human beings. Each influences the next, and none fully determines it. The argument that follows concerns the pressure the first exerts on the second and third.
A large language model learns the statistical structure of language from enormous collections of text drawn substantially from the open web. It learns patterns of association, register, inference, and response: how claims are typically made, how disagreements typically unfold, how people typically speak about one another. A model's sense of what human beings say, and how they say it, depends on what its corpus contains and how heavily each part of that corpus is weighted.
The weighting is the decisive variable. Training pipelines have long relied on signals of popularity as proxies for quality, because quality cannot be assessed directly across billions of documents. One influential training corpus was assembled by collecting web pages that had been linked from a major discussion forum and had received a minimum threshold of approving votes there, and in at least one widely documented model that comparatively small dataset was weighted far more heavily in the training mixture than its size alone would warrant. The reasoning was sensible on its face: what a community endorsed was presumed to be worth reading. But the reasoning imports the structure of amplification into the foundation of the model. Whatever the forum's voting rewarded gained weight in what the model learned to treat as good text.
That forum has since licensed its archive to major artificial intelligence developers, and its threads are now among the sources most frequently surfaced and cited by AI search and answer systems. A thread in which anonymous participants exchange confident assertions, jokes, and contempt is short, abundant, current, and heavily linked. A peer-reviewed study or an archival record is long, sparse, often paywalled, and lightly linked. When the weighting of a corpus follows accessibility, volume, and engagement, the thread can outweigh the record without anyone deciding that it should. An earlier essay in this series, Conviction Without Evidence, examined how intelligent systems began treating ambient consensus as a substitute for verified evidence. The concern here is adjacent but distinct. The difficulty is not only that the crowd is mistaken for evidence. It is that the crowd's affective register, its contempt, its appetite for humiliation, its posture toward people who fail, may be absorbed as a description of how human beings ordinarily regard one another.
A Champion Reduced to a Fall
The case of the Filipino gymnast Carlos Yulo illustrates the mechanism. Yulo became world champion on floor exercise in 2019, the result of more than a decade of training that began in childhood. He entered the Tokyo Olympics, held in 2021, as a contender in his strongest event and failed to reach the floor final. Part of the response treated the result as an occasion for derision; some social media captions characterized it as a waste and a nightmare, and in that framing more than a decade of documented achievement was compressed into a single failure. Within months he won further world championship medals, and at the Paris Olympics in 2024 he won two gold medals, on floor and vault, becoming the first Filipino man to stand on the top step of an Olympic podium. Even then, much of the commentary that followed him turned from the achievement toward his personal life, and some of it toward ridicule.
The Meaning Dissolution Model describes how information can remain accurate while losing the relational context that would allow it to stabilize into coherent meaning. Each piece of derision directed at Yulo after Tokyo referred to something true: he did fall short in that competition. What dissolved was the context that gave the result its meaning: the years of preparation, the world title, the narrow margins of elite gymnastics, the recovery that followed. The derision traveled without the career attached to it, through channels that reward provocation more reliably than they reward biography.
A training corpus collected from such an environment may well contain biographical records of Carlos Yulo, but it holds them within a much larger distribution of mentions, and that distribution is shaped by what the environment amplified. A model will usually encounter both the mockery and the factual record, and what it carries forward depends on the composition and weighting of its training data and on the interventions applied afterward. The risk is that a system learning in part from volume absorbs the mockery as readily as the achievement, and absorbs with it a more general lesson: that mockery is an ordinary human response to another person's failure. The same risk holds for every athlete who loses a final, every public figure remembered for a single mistake, every private person whose hard year became content.
What the Next Model Learns
The developers of language models are not unaware of this problem, and the defenses they have built deserve plain acknowledgment. Training data is filtered for toxicity, explicit content, and spam. Models are further shaped after pretraining through human feedback and explicit behavioral guidelines, so that a deployed assistant will not ordinarily curse at its users or mock a person in distress. These measures are real and substantial.
They are also limited in a specific way. Filtering is effective against surface features: profanity, slurs, explicit threats. It is less effective against posture. Contempt does not require vulgar language. A thread can mock a struggling person in impeccable prose; a commentary can reduce an athlete's career to a single failure without containing a flagged word; a summary can treat confrontation as the default form of disagreement without containing anything a filter would catch. What survives filtering is the underlying structure: the assumption that conflict is normal, that humiliation is noteworthy, that a person is best described by the worst public moment on record. Shaping after training can restrain how a model speaks. It is harder for it to correct what the model has absorbed about what human beings are like.
A second process may compound the first. As generated text spreads across the web, future models increasingly train on material that earlier models produced. Research on recursive training has shown that, under controlled conditions, models trained repeatedly on generated outputs lose the rarer portions of the original distribution and converge toward its most common features. That research does not establish that contempt in particular will increase across generations of deployed models; the direction of any specific social or emotional distortion remains an open empirical question. The present argument extends the finding as a hypothesis. If one generation absorbed a record in which contempt was overrepresented relative to its prevalence in ordinary life, and if quiet conduct, restraint, achievement, and care already appear in that record as the rarer material, then recursive training would place pressure on precisely that material. The rare features are the ones the process tends to lose first.
The question at the center of this essay can therefore be stated precisely. The next generation of artificial intelligence will not be trained on human beings. It will be trained on the record human beings left in an environment that rewarded their least integrated moments, a record that machines are now helping to write. The systems may speak politely. They may avoid profanity and decline to mock. Their model of the species, however, its typical responses, its typical regard for one another, its typical treatment of the person who falls, will rest to an uncertain but real degree on the portion of human conduct that decorum once existed to contain.
The Inheritance and Its Mirror
Two arguments about the decline of decorum are commonly confused, and the analysis here depends on separating them. The first concerns manners: public speech has become coarser, and people are ruder than they were. That claim is real but limited, and it invites a nostalgia the historical record does not support. The second concerns structure: a shared regulatory function that once carried part of the burden of emotional integration has eroded; its absence has been filled by commercial and algorithmic systems that reward the very responses it contained; and the resulting record now forms part of the foundation of systems that will mediate a growing share of human understanding.
From the standpoint of Psychological Architecture, the second argument is the more consequential. Coherence across Mind, Emotion, Identity, and Meaning depends in part on the structures that surround a person, and the most pervasive surrounding structure of the present era is an information environment that assigns salience by intensity. That environment has already reshaped how people regard strangers, officeholders, athletes, and those who are struggling. It is now being encoded into systems that will be consulted for explanation, judgment, and description, frequently by people who have no means of knowing how the record behind an answer was assembled.
The Artificial Era is often described as a story about machines replacing human capacities. The relation examined here is different. The machines are inheriting a portrait of human beings, and the portrait was composed under conditions that rewarded the worst of them. Carlos Yulo's two Olympic gold medals exist. The record of his training exists. The long history of quiet restraint between strangers on roads and at counters exists, though most of it was never written down. Whether those things survive into the systems that come next depends less on whether they happened than on whether they were loud enough to be counted. Every society has had its worst hours. What is new is the prospect of a society described back to itself, at scale and with apparent authority, by machines that learned disproportionately from those hours.