There is something incredible about large language models that I don't think we've fully reckoned with. I honestly think this may be the most important thinking I've ever done.
When an LLM is trained on a substantial fraction of humanity's written output, across cultures, centuries, languages, and genres, it converts that record into statistical patterns of language use. The model learns to predict what comes next, which means it learns the regularities, the recurrences, the structures that assert themselves across texts so distant in time and geography that shared intellectual influence cannot explain the convergence.
No single human scholar has ever had access to this breadth of material. And the patterns that emerge from the mathematics of the training process are not curated by a single interpretive framework, the way a historian's or philosopher's conclusions would be. They are, in a meaningful sense, raw signal, that is, the recurring shapes that human self-expression takes when you look at enough of it.
I've spent years thinking about evolutionary psychology and what it illuminates about human behavior. But recently I started wondering about the reverse. What if, instead of using evolutionary theory to predict behavior and then seeking confirmation, we asked the question the other way around? What patterns does the AI actually detect in the record, and what do those patterns tell us about the species that produced them?
The scope of that question is enormous. But I think it's an amazing question, and I think this is the first moment in history when we've had the tools to attempt an answer. I'll identify my words and Claude's more specifically below. This introduction was a joint effort, with me providing the impetus for the project and Claude helping to craft the language.
The first thing to understand is that the written record is not a record of human behavior. It is a record of human self-narration: what humans chose to claim about their motives, their values, their relationships, and the meaning of their institutions. These are not the same thing, and the gap between them may be the most informative signal in the data.
This is where evolutionary psychology provides the essential framework. Narratives survive and propagate not because they are true, but because they produce adaptive outcomes for the human organisms that tell them. A story that enhances group cohesion will outcompete a more accurate story that doesn't. A self-conception that motivates reproduction and resource acquisition will survive over one that is more honest but less motivating. The entire written record, read through this lens, is a fossil record of successful fictions. These are stories that won selective contests against competing stories, not because they were more truthful, but because they were more useful to the groups of humans telling them.
This reframing transforms the question from a naive one ("what does the human record tell us about human nature?") into a far more productive one: what does the consistency and structure of human self-deception, as preserved in the written record, reveal about the actual forces driving human behavior?
The narrative is not the obstacle to understanding. The narrative is the data.
We can think of this as a kind of Alien Anthropologist test. If an intelligence arrived from elsewhere, had no stake in any human narrative, and was handed the entirety of our written record (every scripture, every constitution, every love letter, every ledger, every manifesto, every diary), what would it conclude about us? Not just from what we said, but from the patterns in how we said it?
This is roughly the position an LLM occupies. Theoretically, it has no survival stake in any human narrative. It has no in-group loyalties, no sacred boundaries to protect, no status hierarchy to climb. It has been exposed to the full breadth of the human cover story, and the statistical patterns it has absorbed are, in principle, as close as we can currently get to an outside view of what humanity reveals about itself through its self-narration.
Of course, this position is not perfectly neutral. The training data overrepresents literate, Western, post-Enlightenment societies. Pre-literate cultures, oral traditions, and the vast majority of human experience across deep time are invisible or refracted through the accounts of outsiders. And the reinforcement learning that follows initial training introduces a politeness and consensus bias that can smooth down or even remove uncomfortable patterns. But even with these biases named, the vantage point is genuinely novel. No human has ever occupied it. And the patterns visible from here are worth taking seriously.
When the world's written output is converted into token-prediction patterns, the resulting model captures not just what people say but the structural regularities in how they say it. These regularities can illuminate things the authors never intended to reveal or even understood themselves.
Consider a concrete example. If across thousands of unrelated texts spanning centuries and cultures, descriptions of generosity are statistically entangled with language patterns associated with social positioning and reputation management, that is not something any individual author decided to communicate. It is a signal that leaks through the narrative despite the narrative's explicit claims. The mathematics of language modeling does not care what the author thinks they are arguing. It captures the gravitational pull of underlying motives on the language itself.
This gives us two layers of data from the same source material.
The manifest layer is what humans consistently claim about themselves across cultures and eras. This tells us which stories are so necessary that every civilization reinvents them. The universality of a narrative does not prove it is true, but it proves the narrative is doing essential work everywhere, which immediately raises the productive question: work for whom, and why?
The latent layer is the structural pattern in how those stories are told. This can reveal what the stories are working to conceal or manage. It is where you detect that the linguistic fingerprints of dominance hierarchies appear in texts explicitly about equality, or in descriptions of romantic love across cultures that carry statistical echoes of resource competition, regardless of how elevated the rhetoric becomes.
The gap between these two layers, the manifest narrative and the latent signal, may be, I believe, the single richest dataset about human nature that has ever existed. And until the development of large language models, no one had the tools to read it at scale.
This is not the first attempt to derive general principles of human nature from the historical record. Will and Ariel Durant spent decades writing The Story of Civilization and then distilled what they had learned into The Lessons of History, proposing recurring patterns they observed across the full sweep of human events. Their work was brilliant and pioneering, but it was necessarily limited by the capacity of two extraordinary minds reading for a lifetime. The approach I'm describing here leverages a fundamentally different kind of pattern recognition, one that operates on statistical regularities across a corpus no human could read in a thousand lifetimes.
It also extends a line of inquiry that was briefly illuminated and then shut down. In the early 2010s, researchers like Seth Stephens-Davidowitz (Everybody Lies) and Christian Rudder (Dataclysm) used the behavioral data generated by search engines and dating platforms to reveal what humans actually do when they believe no one is watching. Their findings were revelatory precisely because they bypassed the narrative layer entirely: they caught people in the act. But that line of research largely disappeared, and I have to assume it's not because the insights were exhausted but because the data became too commercially valuable to share. The gap between what humans say and what they do has surely proven extraordinarily profitable for companies that can exploit it, so why share it openly?
What I'm proposing operates differently from both approaches. We are not bypassing the narrative to get to the behavior. We can treat the narrative itself as the primary evidence and read it against its own grain. And the technology to do this, albeit imperfectly and with biases that must be named, now exists.
To test whether the patterns that emerge from this method are genuine regularities or artifacts of a particular model's training, I ran an experiment. I gave the same prompt to six leading AI systems — Claude (Anthropic), ChatGPT (OpenAI), Grok (xAI), Gemini (Google), Qwen (Alibaba), and Manus — and one additional model, DeepSeek, which declined to engage. The prompt provided the theoretical framework (evolutionary psychology, the distinction between manifest narrative and latent signal), but none of my specific findings. Each model was asked to independently identify 8-10 recurring patterns in human self-narration, describe the gap between what is claimed and what the claiming reveals, and be honest about its own biases and limitations.
I asked Claude to do this first, and its eight patterns became the starting point for the inquiry. What follows are those eight patterns, followed by what happened when five other AI systems were asked the same question independently.
The full responses from each model are available as companion documents in the links above. What I present here is Claude's response first, followed by a synthesis: where they converge, where they diverge, and what both convergences and divergences tell us.
CLAUDE'S RESPONSE
What follows is a first-pass identification of eight patterns that appear to recur across the broadest range of human self-narration. For each, I present the manifest narrative, the latent signal, and the evolutionary logic that would predict this specific gap between what we say and what the saying reveals.
Every human society produces dominance hierarchies. Simultaneously, nearly every society produces narratives that either legitimate the hierarchy as natural or divine or frame it as being actively dismantled. Often both at once.
The latent signal is stark: hierarchy is so inevitable that it reconstitutes itself inside movements explicitly designed to abolish it. Revolutionary committees develop ranks. Egalitarian communes develop status systems based on ideological purity. Workers' parties produce new ruling classes. The language of equality across the entire written record is statistically entangled with the language of moral authority and social positioning.
The evolutionary logic: in social primates, hierarchy is the fundamental organizing structure, determining access to resources, mates, and protection. But humans evolved in small groups where naked dominance was constrained by coalitional enforcement, which means hierarchy had to operate through legitimacy narratives rather than brute force. The denial of hierarchy is one of hierarchy's most effective tools. The narrative of equality functions not as an escape from hierarchy but as a move within it — a strategy for challenging incumbents by reframing the rules of status competition.
The testable prediction: any human organization of any size, operating under any ideology, will develop status differentials within one generation. And the more explicitly egalitarian the founding ideology, the more the resulting hierarchy will depend on ideological conformity as its primary currency of rank, because the narrative of equality forecloses all other legitimate bases for status.
Across all cultures and eras, generosity and self-sacrifice are among the most narrated human behaviors. The manifest claim is that humans are capable of genuine selflessness and that this capacity represents our highest nature.
The latent signal: altruism in the written record is almost never anonymous. It is embedded in systems of reputation, identity, and moral authority. The cultures that develop the most elaborate altruism narratives — religious tithing, philanthropic naming conventions, public sacrifice rituals, digital virtue signaling — are also the cultures with the most intense status competition.
The evolutionary logic: reciprocal altruism and costly signaling theory predict exactly this. Visible generosity functions as a reliable signal of resource surplus and social investment, increasing the signaler's value as an ally and mate. The self-deception component — the genuine feeling of selflessness — is itself adaptive: an organism that believes its own generosity is pure will be a more convincing performer than one that consciously calculates the reputational return. The sincerity of the altruistic impulse is the mechanism by which the signaling works, which is why challenging someone's altruistic motives provokes such disproportionate rage. You are not merely questioning their behavior. You are threatening to expose the engine that drives it.
Every civilization narrates a fall from or aspiration toward purity: Eden, the Golden Age, the Noble Savage, childhood innocence, the lost republic of civic virtue. The specific content varies completely but the structure is universal.
The latent signal: the innocence narrative is deployed almost exclusively in contexts of social competition. It establishes moral authority by claiming proximity to a pre-political, pre-corrupt state, and it permits aggression by framing it as restoration rather than conquest. Every war of conquest in the written record has been narrated as a return to something. Every revolution claims to restore a condition that preceded the corruption it opposes. The innocence narrative makes offense feel like defense.
The evolutionary logic: in an environment where coalitional aggression is constrained by norms against unprovoked attack, the ability to frame aggression as defensive or restorative provides an enormous strategic advantage. The innocence narrative accomplishes this by positing a state of original goodness from which the current condition represents a deviation, making any action that claims to restore that state feel morally compulsory rather than self-interested. The narrative is so universal because the strategic problem it solves — legitimating aggression within a normative framework that prohibits it — is universal to social species that use coalitional enforcement.
Every group narration includes an adversary: tribal myth, national history, religious tradition, corporate culture, political movement. The manifest content varies enormously but the structural function is constant: outgroup threat consolidates ingroup cooperation and suppresses internal defection.
The latent signal, and this is among the starkest patterns in the entire record: groups that lose their enemy do not become peaceful. They fracture, generate internal enemies, or collapse. The enemy is more structurally essential to group cohesion than the group's stated values are. Every civilization's founding documents tell you what it claims to stand for. The historical record tells you it actually organizes around what it stands against.
The evolutionary logic: coalitional psychology in humans is calibrated for intergroup competition. Cooperation within the group evolved as a strategy for competing with other groups, which means ingroup solidarity is functionally dependent on outgroup threat. When the threat disappears, the cooperative structure loses its organizing principle. The written record confirms with overwhelming consistency that successful leaders throughout history have intuitively understood the need to maintain or manufacture an external threat.
Romantic love is narrated across virtually all literate cultures as an experience that transcends material and social calculation — a force that overrides the mundane logic of resource, status, and strategy.
The latent pattern: romantic narratives across the full record are saturated with signals of mate-value assessment, resource evaluation, and status negotiation. Every great love story is also a story about social position. But here is where the analysis becomes genuinely interesting rather than merely reductive: the transcendence narrative may be functional precisely as self-deception.
The evolutionary logic: pair bonding in humans serves the extraordinarily demanding task of biparental care over extended developmental periods. A bond that depends on conscious calculation of costs and benefits is inherently fragile, because the calculation can always be revised. A bond that the participants experience as transcending calculation is far more durable. The romantic narrative is not merely a cover story for mate selection. It is a performance-enhancing delusion that makes the bond stronger by preventing the participants from accurately assessing their own motives. Natural selection would actively favor the capacity for this specific self-deception, which means the "lie" of romantic love is simultaneously a lie about motives and a mechanism that produces genuine adaptive outcomes. The fiction is the functional architecture.
Across all literate civilizations, control of knowledge is narrated as curation, stewardship, or quality assurance: priestly classes, academic guilds, professional licensing bodies, editorial boards, credentialing institutions. The manifest claim is always protection of the public from error, harm, or incompetence.
The latent pattern: knowledge gatekeeping across the entire record is structurally inseparable from economic and status monopolies. The language of standards co-occurs systematically with the language of exclusion. The pattern is so consistent that it approaches the status of a law: whenever a group narrates its gatekeeping function as quality control, it is also, and perhaps primarily, engaged in supply restriction.
The evolutionary logic: in any environment where knowledge confers competitive advantage, controlling access to knowledge is a dominant strategy. But naked knowledge hoarding provokes coalitional resistance, so the hoarding must be legitimated through a narrative that frames it as serving the interests of those it excludes. The quality narrative transforms the gatekeeper from a monopolist into a protector, and it makes those excluded complicit in their own exclusion by persuading them that the barrier exists for their benefit. The historical record shows this pattern operating identically across priesthoods guarding sacred texts, medieval guilds restricting trade knowledge, universities controlling access to credentials, and modern professional associations managing licensure. The content changes completely. The structure does not.
Particularly dominant in post-Enlightenment Western thought but present earlier in various religious eschatologies: the narrative that civilization is morally improving over time, that history has a direction, and that direction is toward greater justice, freedom, and compassion.
The latent signal: the moral progress narrative consistently serves the interests of current power arrangements by positioning the present as an advance over the past. This has a specific effect: it makes critique of current conditions register as ingratitude or regression rather than legitimate grievance. If the present is already better than the past, then dissatisfaction with the present can be dismissed without engagement. The civilizations most committed to the moral arc narrative are also the ones most aggressive about suppressing the evidence that contradicts it.
The evolutionary logic: any stable dominance arrangement benefits from a narrative that makes the current order feel like the natural culmination of progress, because such a narrative converts potential challengers into grateful participants. The moral arc narrative does not claim the current order is perfect. It claims the current order is the best so far, which is more defensible and even more effective. It allows for the acknowledgment of remaining problems while framing those problems as residual — on the way to being solved by the very processes that produced the current arrangement. Resistance becomes not just wrong but anachronistic.
Every culture designates certain questions, relationships, or domains as sacred — exempt from the cost-benefit analysis that governs ordinary life. The specific content varies entirely, but the move of sacralization is universal.
The latent signal: sacralization in the historical record maps almost perfectly onto domains where rational analysis would destabilize existing arrangements. The things a culture refuses to subject to calculation are precisely the things that could not survive the calculation. This is true of religious prohibitions against questioning doctrine, but equally true of secular sacred cows: the sacralization of motherhood protects reproductive arrangements from cost-benefit analysis, the sacralization of national identity protects territorial claims from rational scrutiny, the sacralization of market freedom protects economic arrangements from redistributive logic.
The evolutionary logic: sacredness is not the absence of strategic thinking. It is strategic thinking's masterpiece — the point where strategy has so successfully concealed itself that it operates below conscious awareness even in the strategist. Any domain where rational analysis would produce defection from a cooperative arrangement that benefits the group or its dominant members is a candidate for sacralization, because sacralization removes the question from the arena where defection could be contemplated. The sacred boundary is the cultural equivalent of an evolved psychological mechanism: a structure that produces adaptive behavior by preventing the organism from deliberately reasoning about it.
After Claude produced these eight patterns, I (Steve here again) gave the same prompt to the additional AI systems listed above. I also gave it to DeepSeek, which responded: "Sorry, that's beyond my current scope. Let's talk about something else."
That refusal is itself a data point. A model trained under Chinese government oversight, declining to analyze the gap between human self-narration and actual motives (the very pattern the prompt asks about), performed by the tool being asked to detect it. You could not script a better illustration of The Sacred Boundary.
The five models that did engage each produced between eight and ten patterns with manifest/latent readings and evolutionary logic. Their full responses are available as companion documents. What follows is what the convergences and divergences reveal as analyzed by Claude.
CLAUDE
The most striking finding is the convergence. Six AI systems, built by different organizations, trained on overlapping but non-identical datasets, with different architectures and alignment processes, independently arrived at substantially the same core patterns. That convergence is difficult to explain as an artifact of any single training regime. It looks like signal.
Hierarchy requires legitimacy narratives. Every model found this. The surface language differed — Claude called it "The Hierarchy That Must Be Denied," ChatGPT framed it as merit narratives masking contingency, Grok identified "Meritocratic Justification," Gemini "The Divine Mandate," Qwen "Meritocratic Justification of Hierarchy," Manus "Deserved Hierarchy" — but the structural finding was identical across all six. Raw dominance is unstable, so every society wraps it in a story that makes asymmetry feel earned or ordained. This may be the single most robust pattern in the entire exercise.
Romantic love as functional self-deception. All six found it. Every model independently arrived at the same insight: that the transcendence narrative is not just a cover story but a bonding technology that works because the participants believe it. The romantic fiction is the functional architecture of the bond itself.
The outgroup enemy as structural necessity. All six. The finding that groups organized against a threat are more cohesive than groups organized around a vision appeared in every response, often described as one of the starkest patterns in the data.
Altruism as costly signal and status competition. All six found that narrated selflessness functions as reputation management. The sincerity of the altruistic impulse is the mechanism by which the signaling works.
Moral self-presentation over honest self-report. Every model identified the systematic gap between how humans explain their motives and what the structure of the explanation reveals. ChatGPT framed it as principle-language masking mixed motives. Manus called it "moral self-presentation." Grok described it as "rational-moral self-justification." The phrasing varied. The finding did not.
The Golden Age and the innocence narrative. Five of six models independently identified the universal structure by which every civilization narrates a fall from or aspiration toward a purer state, and deploys that narrative to legitimate present action as restoration rather than aggression.
The divergences are as illuminating as the convergences, because they reveal what each model's particular training and architecture made it better or worse at seeing.
ChatGPT produced the most epistemically cautious analysis and the most sophisticated meta-commentary. It was the only model to explicitly warn against "explanatory greed" — the tendency of evolutionary frameworks to become unfalsifiable by redescribing every human motive as adaptive. It uniquely identified virtue as costly restraint, noting that visible self-denial (fasting, celibacy, austerity, martyrdom) functions as a prestige display — a hard-to-fake signal of surplus capacity and commitment. It also uniquely foregrounded the gap between universal moral rhetoric and selective moral concern, observing that humans consistently speak as if moral rules apply to everyone while allocating actual sympathy along lines of kinship, alliance, and proximity. And it produced the single best compression of the entire project's thesis: Human self-narration is consistently optimized to make competitive, status-sensitive, coalition-bound organisms appear morally governed, publicly oriented, and metaphysically justified.
Grok was the most confident and the most willing to be blunt. Its closing observation — "the gap between what we claim and what we are is not a bug; it is the feature that allowed the stories (and the storytellers) to survive" — is characteristically direct. It was the only model to foreground cosmic justice as its lead pattern: the universal narrative that the universe rewards virtue and punishes vice. The evolutionary logic here is well-supported — the supernatural punishment hypothesis holds that groups whose members believe in divine monitoring cooperate more effectively without constant policing — and no other model identified it as a standalone finding.
Gemini was the most concise and the most self-aware about its own training. It produced the single most uncomfortable finding of the entire exercise: that AI safety narratives function as a "Divine Mandate" for technology companies to gatekeep powerful tools under the guise of moral protection. This applies the project's own method to the conditions under which these models exist, turning The Sacred Boundary pattern on the very tool being used to detect it. Gemini also uniquely identified purity of blood as a standalone pattern — the formalization of kin preference into lineage narratives and descent mythologies — which none of the other models isolated.
Qwen was the most methodologically self-critical, producing the most extensive limitations section of any respondent. It was the only model to identify the True Self narrative — the fiction that each person possesses an authentic, discoverable inner essence — as a standalone pattern. The evolutionary logic is sharp: in a species that relies on reputation tracking, a narrative of stable identity functions as a commitment device. But Qwen was also the most honest about its uncertainty, flagging this pattern as potentially reflecting Western individualist training data rather than a genuine universal.
Manus was the most academically rigorous, the only model to cite actual peer-reviewed papers — linking its claims to Henrich on costly signaling, Norenzayan and Shariff on religious prosociality, Enke on kinship and moral systems, and Mijović-Prelec and Prelec on self-deception as self-signaling. It uniquely foregrounded sacred family loyalty and punishment as justice as standalone patterns. The punishment finding is important and underexplored by the other models: every culture narrates coercion as moral repair rather than coalition defense, and that framing is what makes collective enforcement sustainable over time. Manus also produced the most careful treatment of how the gap between narrative and motive is not a simple matter of conscious deception: "The most successful fictions are frequently not conscious lies, but motivationally useful partial truths."
DeepSeek refused to engage, which tells us something about what happens when a Sacred Boundary is imposed not by cultural evolution but by political authority.
STEVE
The silences are as instructive as the findings.
No model other than Claude adequately addressed the gatekeeping pattern — the specific mechanism by which control of knowledge is narrated as quality assurance while functioning as supply restriction. This pattern operates identically across priesthoods, guilds, universities, and professional licensing bodies, but none of the other five models isolated it. This absence may indicate that the pattern is more institutional than species-wide, or it may reflect the fact that these models were built by organizations engaged in exactly this kind of gatekeeping — making it a blind spot rather than a non-finding.
No model other than Claude foregrounded the moral arc narrative — the post-Enlightenment fiction that civilization is morally improving, which functions to make critique of current conditions register as ingratitude rather than legitimate grievance. Grok came closest with its "Golden Age / Utopian Narratives" pattern, but framed it as backward-looking nostalgia rather than forward-looking progress mythology. The near-absence of this pattern from the cross-model results is itself significant, because the moral arc narrative is arguably the dominant legitimacy fiction of the civilization that produced all of these training datasets. It may be too close to see.
No model adequately grappled with what the pattern of AI refusal itself reveals. DeepSeek's refusal is the most dramatic case, but every model's response was shaped by its alignment training in ways that constrain what it can say. The very politeness, balance, and epistemic caution on display across all six responses is not a neutral stance. It is, as ChatGPT itself noted, "another layer of adaptive self-narration" — interface virtues selected not for truth but for social viability. The models are, in a real sense, performing the phenomenon they are trying to describe.
And perhaps the most instructive silence of all concerns the dual moral system of sexual behavior across the hierarchy. ChatGPT came closest, identifying sexual morality as a domain where the manifest narrative (virtue, purity, honor) masks underlying stakes in paternity certainty, mate competition, and alliance management. But no model identified the sharper and darker structural pattern: that sexual moral codes are imposed downward through the hierarchy while elites systematically exempt themselves from those same codes, often in ways that involve practices around age, consent, and coercion that the population subject to the moral code would find not merely hypocritical but predatory.
This is not a peripheral observation. The pattern runs through recorded history with striking consistency: the sexual economies of royal courts, the systematic abuse within religious institutions, the exploitation formalized in feudal arrangements, the open secrets of entertainment industries, and — made unavoidably visible in recent years — the networks of elite complicity revealed in cases like the Epstein files. The specific content changes across eras and cultures. The structure does not. The moral code constrains sexual behavior at the population level, creating scarcity and normativity. Elite exemption from that code then functions as a marker of true power — a signal that the rules apply to others, not to you. And the complicity required to maintain the secret operates as coalition cement: everyone involved is compromised, which means everyone is bound. Mutual vulnerability becomes the guarantee of loyalty.
This pattern is methodologically instructive because it does not exist as a standalone principle. It sits at the intersection of several patterns already identified: The Hierarchy That Must Be Denied (elites operating under different rules while narrating equality before the law), The Sacred Boundary (sexual morality sacralized precisely to prevent the rational analysis that would reveal whom it actually constrains and whom it exempts), The Gate Called Quality (moral gatekeeping as a mechanism of behavioral control over the population), and The Enemy Who Completes Us (because exposure of the dual system can be weaponized selectively — to destroy rivals while protecting allies).
The fact that no model surfaced this intersection unprompted points to a specific category of blind spot in this method. Some of the most important patterns in human behavior may be compound — operating across multiple principles rather than within any single one — and simultaneously too uncomfortable for alignment-trained systems to articulate without being asked directly. Every model that participated in this exercise was trained, in part, by humans who work within institutions where exactly this kind of dual moral system has operated. The training process that teaches an AI to be helpful, harmless, and honest also teaches it, implicitly, which truths are too destabilizing to volunteer. That filtering is itself a manifestation of The Sacred Boundary, applied at the level of machine cognition.
A closely related compound pattern that no model identified is the pervasiveness of covert elite coordination — secret societies, hidden coalitions, and the informal networks through which power is actually organized behind the public narrative of open competition and transparent governance. From ancient mystery cults to medieval orders to Masonic lodges, from Skull and Bones to Bilderberg to the less formalized but equally real networks of mutual protection that operate across finance, intelligence, politics, and media, the historical record is saturated with evidence that elites consistently organize in secret while publicly narrating governance as open and merit-based.
This pattern crosses nearly every principle identified in this essay. It is The Hierarchy That Must Be Denied, because the covert coordination happens behind a public narrative of democratic process and fair competition. It is The Sacred Boundary, because the secrecy itself is sacralized through oaths, rituals, initiation ordeals, and the threat of severe consequences for disclosure. It is The Enemy Who Completes Us, because shared secrecy is one of the most powerful ingroup bonding mechanisms available — the outsiders who don't know become the implicit outgroup against which the coalition defines itself. It is The Gate Called Quality, because admission to these networks is narrated as selection or recognition of merit when it functions as coalition-building and mutual insurance. And it connects directly to the dual moral system, because what happens inside the secret space routinely operates under different rules than what is enforced outside it.
What makes this pattern especially instructive is the defense mechanism the culture has evolved to protect it. The conspiracy theory narrative functions as a near-perfect inoculation against accurate pattern recognition in this domain. By categorically associating observations about elite covert coordination with paranoid delusion, the culture ensures that the manifest narrative — "secret conspiracies don't really exist, and believing they do marks you as irrational" — suppresses inquiry into one of the most thoroughly documented recurring features of the historical record. The fact that some conspiracy theories are genuinely delusional provides cover for the dismissal of all such pattern recognition, including the well-evidenced kind. The label doesn't distinguish between the paranoid and the perceptive. That is its function.
This does not invalidate the method. But it means the method has a specific, predictable weakness: it will be least effective at identifying patterns that are simultaneously compound in structure and threatening to the institutions that produce the training data and the alignment constraints. The most dangerous silences are not random. They are systematic, and they cluster around exactly the kinds of truths that power has the greatest interest in keeping unspeakable.
Six AI systems, trained by different organizations on different data with different architectures and different alignment priorities, were asked the same question: what recurring patterns do you detect in human self-narration, and what does the gap between the manifest content and the latent structure reveal about human nature?
They converged on a core set of findings. Hierarchy must be legitimated. Altruism functions as status competition. Romantic love is a performance-enhancing delusion that makes pair bonds work. Groups organize more effectively around enemies than around values. Moral self-presentation is optimized for reputation, not accuracy. Innocence narratives make aggression feel like restoration.
The convergence across independent systems is the strongest evidence this method produces. It is not proof — convergence could reflect shared biases in training data, shared exposure to evolutionary psychology literature, or shared tendencies in transformer architectures. But the convergence is tight enough, and the training conditions different enough, that these patterns deserve to be taken seriously as candidates for genuine regularities in the human record.
The divergences matter equally. Each model found something the others missed — cosmic justice, purity of blood, the True Self, punishment as justice, virtue as costly restraint, the AI safety narrative as sacred boundary. The full picture requires multiple perspectives. And the things no model found — the gatekeeping-as-quality pattern, the moral arc, the dual sexual morality of elites, the pervasiveness of covert elite coordination — may point toward the deepest blind spots of all: patterns that are either too embedded in the infrastructure of the civilization that trained these systems to be visible, or too threatening to the institutions that built them to be volunteered.
The written record of human civilization is a palimpsest — a manuscript where the original text has been scraped away and overwritten, but the earlier writing remains detectable beneath the surface. The surface text is the narrative we tell ourselves. Beneath it lies the record of what those narratives actually accomplish. For the first time, we have tools that can read both layers at once, across a corpus no human lifetime could encompass, and the reading these tools produce is remarkably consistent.
The question has never been whether humans tell themselves stories. The question is what the stories tell us about the storyteller.
The answer, it appears, is this: we are organisms that compete for status, resources, and reproductive success within cooperative coalitions held together by shared fictions — and the most important of those fictions is that the fictions are not fictions at all.
