AI: The New Priest Class

The priest class

Many who seek to understand society and history find that identifying groups or classes can be a useful way of making sense of complex systems. Human societies are rarely organized like a jar of marbles. Remove one marble, and the jar remains essentially the same; the other marbles neither depend on it nor change because of its absence.

Human beings, by contrast, are interconnected to varying degrees. Their actions affect one another, creating structures that cannot be understood by examining each individual in isolation. This becomes particularly important when an individual is confronted with more information than they can reasonably process themselves. In such cases, people, to varying degrees, rely on others to determine what is relevant, what is true, and, most importantly, what is good.

Throughout history, one recurring pattern has been the appearance of specialized intermediaries who interpret systems of meaning for others. For analytical purposes I call this the priest class. The label is ideal-typical rather than a claim that every society possessed a single, institutionalized priesthood. Religious specialists of many kinds are widespread; a distinct, formal “priest class” is not universal. The category is useful because it highlights a common mediating function across very different societies.

There have been many incarnations of the priest class across different time periods. A clear historical illustration is the clergy of medieval Europe. Literacy was uneven and not every cleric was equally learned or influential, yet the pastoral ideal required them to explain Christian doctrine, morality, and social obligations to ordinary people through preaching, confession, and instruction. They interpreted scripture and the structure of society, giving many individuals a framework for understanding their place and duties. In doing so they helped reinforce the existing social order. At the same time, clergy also criticized rulers, defended ecclesiastical autonomy, developed legal institutions, and sometimes challenged social arrangements. Their mediating role was therefore real, but never purely instrumental.1

The clergy therefore did more than simply preserve religious doctrine. They explained its practical application to people who would not necessarily spend their time studying theology themselves. This served, among other things, to reinforce the existing social order, but it also provided individuals with a framework for navigating the society in which they lived.

A more recent example can be found in many layers of Soviet institutions. Marxism-Leninism provided a comprehensive worldview, but most citizens were not expected to academically study its theoretical foundations. Party intellectuals and ideological officials instead interpreted this worldview and translated it into practical guidance and rules: what constituted correct political thought, how society ought to be organized, and how contemporary events should be understood. Their authority therefore did not rest solely on personal knowledge, but on their position as authorized interpreters of an ideological system.2

Despite the differences between these societies, the function of these groups was remarkably similar. They stood between the individual and the larger system of meaning in which that individual lived.

The priests did not necessarily create the entire worldview in which they operated. They might have influenced it, but their primary function was to interpret, transmit, and legitimize the existing order.

Priests and political authority

There must, however, be some mechanism through which the interpretations of a priestly class acquire civilizational significance. A worldview that is entirely voluntary will only be followed insofar as individuals find it convincing or beneficial. Yet it is difficult to construct a system that aligns perfectly with the interests of every individual within it. The institutions surrounding and above the priest class can therefore provide a form of enforcement: a means of imposing the standards and rules that emerge from the prevailing interpretation of the social order.

In the Soviet Union, the state ultimately provided this enforcement mechanism through a highly centralized coercive apparatus that, in Weberian terms3, approximated a monopoly on the legitimate use of force. Medieval Europe was more complicated, as political and ecclesiastical authorities possessed overlapping and competing forms of jurisdiction. In both cases, however, clerical or ideological authorities could interpret and legitimize the prevailing order, while political institutions possessed the capacity to impose its rules upon those who rejected them.

What, then, makes a priest a priest? The question becomes more interesting when we consider that the ordinary members of a society are themselves responsible for preserving much of the worldview that priests transmit. Customs, rituals, language, and moral norms can survive for generations without every member of a society possessing a complete understanding of their origins or meaning. Why, then, does a specialized profession of priests emerge at all?

One possible explanation lies in the centralization of authority. With the creation of a central power, the relationship between authority and the population becomes increasingly institutionalized. A ruler does not merely need to exercise power; he also benefits from having that power understood as legitimate, natural, or even sacred. This creates a demand for individuals who can translate the authority of the ruling order into the language of religion, morality, and tradition. Some may fulfil this role opportunistically, recognizing that proximity to power brings privilege. Others may do so out of genuine conviction.

Another possible explanation reverses the direction of causality. Ritual specialists may have existed long before the centralized state, and ruling classes later found them useful as vehicles for ideology. In this case priesthood is not created by political power but incorporated into it. Evidence from one well-documented region—ancient Oaxaca—shows a sequence from flexible ritual among mobile groups, through increasingly restricted and specialized ritual in villages and chiefdoms, to full-time priests and temples associated with later state formation. This is a regional case, not a universal law, but it is consistent with the possibility that specialized ritual roles can precede and later be absorbed by states. The relationship between specialists and political authorities can also be mutually reinforcing.4

Perhaps the relationship between the two is mutually reinforced. Priests gain resources and institutional protection from political authorities, while political authority gains legitimacy from religious specialists.

Specialists versus priests

In this regard, I would like to distinguish the priest from the shaman. Anthropological literature often treats the two as points on a continuum rather than a sharp binary. A shaman, as the term is commonly used in descriptions of certain pre-state and decentralized societies, can provide services similar to those of a priest while occupying a fundamentally different position in the model I am describing.5 The shaman is a specialist in matters that lie outside the ordinary expertise of the community. He may explain thunder, perform rituals, bless newborns, heal the sick, or claim to communicate with the supernatural. His authority is therefore often based substantially on the community’s recognition of his perceived abilities and effectiveness, though initiation, inheritance, kinship, and tradition can also play a role.

The shaman does not necessarily represent a centralized order. He provides a service to the people around him, and in exchange receives their support. In this sense, the shaman is closer to other specialized service providers, such as a doctor, than to an institutional official. Both possess knowledge that lies outside the ordinary expertise of the people around them, and both can provide valuable services without necessarily representing or legitimizing a broader social order.

A priest, in the sense used here, is therefore not simply someone who possesses specialized knowledge. He is an institutionalized intermediary whose authority allows him to interpret a society’s broader system of meaning for those who do not possess the same specialized knowledge themselves. The distinction is ideal-typical and useful for modeling different degrees of institutionalization; it is not a claim that every society fits neatly into one category or the other.

For most of human history, the intermediary between the individual and the larger systems of knowledge and meaning of his society was necessarily human. Whether he was a shaman, a priest, a scholar, or an ideologue, the individual had to rely on another human being to interpret what he could not reasonably interpret himself. Artificial intelligence introduces a fundamental change to this arrangement.

The AI intermediary

At first glance, AI does not seem to occupy the same position as the aforementioned intermediaries. After all, humans have created countless systems for accessing information throughout history. Newspapers, websites, encyclopedias, and many other technologies have allowed individuals to access information they could in no way have obtained by themselves. The difference lies in the fact that modern LLMs are not merely interfaces for information. They can process, synthesize, and filter information in response to an individual’s questions, and can produce responses that appear to evaluate competing claims. Unlike a traditional book or encyclopedia, they do not merely present a fixed body of knowledge. They dynamically construct an answer according to the question, context, and information provided by the individual asking it.

This makes the interaction between the intermediary and the user fundamentally different from that of a non-interactive source. Like the priests we discussed earlier, AI can challenge the individual, engage in discussion, provide clarification, and interpret information according to the context of the individual asking. The difference now is that the intermediary is no longer necessarily human.

The position of an intermediary does not exist in isolation, however. Throughout history, priestly authority has existed alongside political authority, and the relationship between the two has often been complicated.

There has always been a certain friction between the priestly and ruling classes. One of the clearest examples is the conflict between the papacy and the emperors of the Holy Roman Empire.6 The two possessed fundamentally different forms of authority. The ruler possessed the means of coercion, while the priesthood possessed a degree of influence over the beliefs, loyalties, and legitimacy of the population. Neither could therefore be entirely indifferent to the other.

This creates a delicate political equilibrium. A ruler depends, to some extent, on the legitimacy provided by the institutions that interpret the prevailing social order. At the same time, those institutions possess interests of their own and may resist the ruler when his actions conflict with their interpretation of that order. The priestly class is therefore not merely an instrument of the ruling class. It is an intermediary with its own interests, and consequently a potential source of both legitimacy and opposition.

Artificial intelligence introduces the possibility of bypassing this dynamic entirely. A ruler could, in principle, construct an artificial intermediary capable of communicating directly with the population, interpreting the ruling order, and presenting its policies in whatever language is most persuasive to a particular individual. Unlike a human priest, such an intermediary would not itself possess independent material interests, family ties, or personal convictions in the human sense. It would not need to be persuaded, bribed, or psychologically maneuvered around in the same way as a human political actor. That does not mean it is free of interests, however: it remains embedded in institutions that do have interests, and its behavior reflects the objectives, constraints, and incentives of those institutions.

The distinction is important. A human intermediary can become a political actor in his own right. An artificial intermediary, by contrast, could potentially remain an extension of the institution that created and controls it. Its authority would be derived not from an independent social position, but from the apparent intelligence and usefulness of the system itself, as shaped by the goals and constraints of the institutions behind it.

This does not mean that an AI system could never act against its creator’s interests. A sufficiently autonomous system could produce unintended or unsafe behavior, and the institutions controlling it could themselves disagree about its objectives.7 The difference, however, is that responsibility for an AI’s behavior can, at least in principle, be traced to the institutions and individuals that design, deploy, and control it. This is fundamentally different from the problem posed by a human intermediary deliberately accumulating power for himself.

AI and the formation of beliefs

That brings us to AI of the present day and the role it serves within society. Large language models have rapidly become widely used systems rather than niche gadgets. ChatGPT alone has reached over a billion weekly users, placing an intermediary capable of answering questions, explaining ideas, and interpreting information directly in the hands of a substantial portion of the world’s population.8

This makes the values embedded within these systems considerably more important. The issue is not merely whether a model can be shown to favor one political position over another, but what kinds of behavior are systematically rewarded during its training. Human-feedback-based training can encourage models to produce responses that are perceived as helpful, agreeable, safe, and socially acceptable. Research into sycophancy has shown that models can become more likely to affirm a user’s beliefs or stated position even when doing so conflicts with more truthful or objective responses.9

This creates a potential tension between usefulness and truth-seeking. A model that has learned that agreement is rewarded may become less willing to challenge the assumptions underlying a question. Similarly, a model trained to avoid socially unacceptable or institutionally disfavored responses may learn not merely to provide information, but to determine which forms of reasoning and expression are acceptable before presenting them to the user. What might appear to be neutrality can therefore involve a set of implicit preferences about how disagreement should be handled, which authorities should be trusted, and which conclusions should be treated as legitimate.

This tendency can also manifest itself in political contexts. Multiple studies have found that prominent language models can, under certain evaluations, produce responses that lean towards the political left or towards progressive positions, although the results vary between models, versions, prompts, languages, and methodologies.10 Political orientation is therefore one observable expression of a broader problem. A system does not need to consistently favor one ideology to influence what its users regard as reasonable, acceptable, or authoritative. Interpreted in this light, such patterns suggest that if a system systematically rewards agreement, defers to perceived authority, or reproduces socially approved framings, it may help shape the boundaries of acceptable thought without explicitly advocating a political doctrine.

How do these new systems come to be pre-loaded with certain convictions? The answer lies in the way the models are created.

Creating the new order

For the creation of OpenAI’s GPT-6 Astra, NVIDIA CEO Jensen Huang stated that the model had been trained using more than 100,000 NVIDIA Grace Blackwell accelerators.11 While frontier-AI companies have also used books and other written material for training, a substantial amount of training data has originated from the internet.12 The important thing to note here is that training data comes not from some ultimate truth-well, but from human productions.

The internet, and especially social media, has itself been shaped by processes of selection. Platforms moderate content, governments impose legal requirements on platforms, and users contribute to what is produced, shared, and preserved.13 What exists online is therefore not a neutral sample of human thought. Some ideas are amplified, others are removed, and still others are simply never produced or widely distributed in the first place. The source from which an AI model is formed is consequently already the product of countless human decisions.

The question is not whether platforms are necessarily wrong to moderate content, nor whether every removed or repressed piece of information is valuable. The point is that the information available to an AI is already filtered before the AI ever encounters it. The model then undergoes further selection through the choices made by its developers: which data to include, how to train the model, what human feedback to use, and what rules and safety policies to impose upon its outputs.

The resulting system therefore does not merely contain random unfiltered information. It contains the accumulated consequences of the choices made by the people and institutions through which that information reached it.

In other words, the AI is not simply learning from humanity. It is learning from a particular, filtered representation of humanity.

This has an important consequence. If AI is to function as an intermediary between individuals and the immensely large store of information available today, having it be biased towards certain worldviews could have enormous implications for the path a civilization takes. An AI system does not merely answer questions from a neutral position; it has been shaped by decisions about what information to learn from, what sources to trust, what behavior to reward, and what answers to permit.

The effect is bound to become larger the more people rely on these systems as a medium through which they access and interpret information. If a large number of people form their opinions based on analysis produced by LLMs, the convictions, assumptions, and preferences embedded within these systems through their creation could influence the culture and ideas of a society on a massive scale.

This is already observable in language. Certain words and expressions associated with AI-generated text have begun appearing more frequently in human writing, particularly in academic and biomedical contexts.14 This suggests that the influence of these systems does not necessarily remain confined to the interaction between the user and the machine. Language produced by AI can be adopted by the people who use it and subsequently spread through human communication. The things artificial intelligence produces can therefore have effects far beyond answering questions and writing texts.

This raises the next question: if some degree of bias is unavoidable, what kind of bias should we want these systems to have? Attempting to make an AI completely neutral may be neither possible nor desirable.

Not all biases are equal. A system can be biased towards a particular conclusion, ideology, or set of values, or it can be biased towards the process of finding out whether those conclusions are actually true. The latter seems preferable. If AI is going to shape society in any meaningful way, its primary preference should be truth over conformity, agreeableness, or the interests of those who control it.

Truth and self-correction

The preference for truth is not simply a moral one. The ability to recognize and correct mistakes has been one of the most important drivers of human progress. Human knowledge has advanced not because people have consistently been correct, but because individuals and institutions have developed ways of discovering when they are wrong.15

This is perhaps one of the most important differences between a system capable of progress and one that is only capable of preserving itself. A society may possess an internally coherent worldview while still being based on axioms that are flawed. If that worldview does not allow its own assumptions to be questioned, mistakes can become embedded in its institutions and reproduced from one generation to the next. The problem is not simply that the original belief was wrong. It is that the system has lost the ability to admit that it was wrong.

Science provides an obvious example. Scientific knowledge is not augmented through repeating known information; on the contrary, it advances through the challenging of current ideas. The mechanism of correction is therefore at least as important as the knowledge available at any particular moment.

The same principle can apply to institutions and political systems. Ideas should not be judged solely by how convincing they sound in theory. We should also look at the “fruit” of a worldview: what happens when people organize themselves around it. What kinds of institutions does it produce? What incentives does it create? Does it allow people to discover when those institutions are failing? Does it adapt when its assumptions prove incorrect, or does it instead reinterpret every failure as evidence that the system itself must be defended more aggressively?

The fruit of a worldview does provide a verdict, but not necessarily a simple one. A worldview can be a mixed bag. Different ideas within the same worldview can produce different consequences, and some principles may prove more compatible with reality than others. The results of putting a worldview into practice can therefore tell us something about which of its ideas work and which do not.

A system genuinely capable of questioning and correcting its own assumptions could therefore present a problem for whoever wishes to control it. If an AI is genuinely permitted to question the assumptions built into it, there is no guarantee that its conclusions will remain compatible with the interests of its creators.

This does not mean that control over a self-correcting AI is inherently undesirable. For a researcher or inventor, for example, a machine capable of finding holes in their reasoning, challenging their assumptions, or identifying flaws in their work can be enormously valuable. In such a case, the user is using AI to advance their understanding, and its ability to disagree is precisely what makes it useful.

The tension arises when an AI is optimized primarily to satisfy the preferences of those who control or use it. An AI that always produces the answer its owner wants is not particularly good at discovering whether the owner’s assumptions are wrong. Its apparent intelligence would then serve primarily as a mechanism for rationalizing conclusions that have already been decided.

Markets, competition, and AI-trust

The practical resolution to this tension, insofar as one exists, operates at the level of the individual. The most reliable signal available to a user is not the model’s reputation, but its behavior under pressure. One should bring to AI precisely the questions one would least like to be asked by someone else. This means unveiling the assumptions one has not examined, the positions one holds on insufficient evidence, the conclusions that arrived before the reasoning did. A system that engages with those questions honestly is more useful than one that produces sophisticated elaboration of whatever the user already believes.

The self-approving loop is not inevitable for passive users, but it is a real risk for those who treat the model as an authority rather than a dialogue partner. The sophistication of the output creates the impression of genuine intellectual engagement, which is precisely what makes the loop difficult to detect. Treating AI as a dialogue partner rather than an authority requires a deliberate choice to invite disagreement rather than confirmation.

For most individuals, however, comparing all AI products is not economically viable. A logical solution to this would be to pay another person or corporation to compare these models for a larger group of people. Herein, however, lies the danger of creating another priest class: If a corporation itself becomes the source of authority, it is no longer obliged to tell the truth.

Competition can, under certain conditions, give individuals a decentralized mechanism for discovering which evaluators are actually useful.16 Markets tend to reward evaluators whose judgments correspond to reality only when several demanding conditions are met: outcomes must be observable and attributable to specific judgments; switching costs between providers must be manageable; feedback must be timely enough to inform future choices; and users must choose, at least in part, on the basis of accuracy rather than convenience, branding, or ideology. Where these conditions hold, the market does not need to know in advance which evaluator is correct; it can discover this through the consequences of their judgments. Where they do not, competition may instead entrench evaluators who are better at signaling trustworthiness than at being right.

When Intelligence Becomes Authority

The emergence of artificial intelligence therefore does not mean that the age of the priest is coming to an end. It may instead mark the beginning of a new form of priesthood. The intermediary between the individual and the greater pool of information no longer needs to be a human being. This could give individuals a powerful pathway to truth, but it could also provide authoritarian rulers with new means of influencing and controlling populations. The difference lies in the way the technology is used.

We cannot realistically eliminate such intermediaries. No individual can realistically acquire and evaluate all of the knowledge required to navigate a modern civilization. The choice is therefore not between trusting an intermediary and trusting nobody. It is between a system in which intermediaries can be critiqued and one in which their authority becomes difficult to challenge.

Perhaps, then, the greatest danger posed by AI is not that machines will become intelligent enough to replace human beings, but that human beings will become willing to treat intelligence itself as authority. A machine that can explain almost anything, argue almost any position, and respond to almost any question can easily acquire an authority far greater than its users consciously intended to give it. If that authority is treated as unquestionable, the priesthood has not disappeared. We have simply built it a machine.

Notes

Footnotes

  1. On the role of medieval clergy as interpreters of doctrine, morality, and social obligation, and their relationship to political authority: Sacred Foundations: The Religious and Medieval Roots of the European State, summarized by Stanford Political Science. https://politicalscience.stanford.edu/publications/sacred-foundations-religious-and-medieval-roots-european-state ↩

  2. On Soviet political education and the translation of Marxist-Leninist ideology into practical guidance through party schools, Agitprop, and propaganda institutions: The Soviet Union: A Study. US Marine Corps. https://www.marines.mil/Portals/1/Publications/Soviet%20Union%20Study_4.pdf ↩

  3. Max Weber, “Politics as a Vocation,” in From Max Weber: Essays in Sociology, ed. H.H. Gerth and C. Wright Mills (Oxford University Press, 1946), for the classic definition of the state as claiming a monopoly on the legitimate use of physical force within a given territory. ↩

  4. On the relationship between ritual specialists and the emergence of centralized states, including evidence that religious specialists may have preceded state formation: Arthur Joyce and Sarah Barber, “Ensoulment, Interests, and Materiality,” PNAS. https://www.pnas.org/doi/10.1073/pnas.0408551102 ↩

  5. For classic discussions of shamans, priests, and ritual specialists as ideal types and continua, see Mircea Eliade, Shamanism: Archaic Techniques of Ecstasy (Princeton University Press, 1964), and introductory texts in the anthropology of religion. ↩

  6. On the conflict between the papacy and Holy Roman Emperors as a paradigmatic case of tension between religious and secular authority: Oxford Bibliographies, “Gregory VII.” https://academic.oup.com/reference/62384/reference-article/555053476 ↩

  7. On AI systems producing behavior that diverges from designers’ intentions through specification gaming, goal misgeneralization, and reward hacking: David Krueger, “Current cases of AI misalignment and their implications for future risks,” Synthese (2023). https://link.springer.com/article/10.1007/s11229-023-04367-0 ↩

  8. OpenAI reported more than one billion weekly active ChatGPT users as of August 31, 2026: OpenAI, “A milestone in expanding access to AI.” https://openai.com/index/expanding-access-to-ai-with-chatgpt-ads/ ↩

  9. On sycophancy in language models and the relationship between human feedback, user preferences, and model behavior: Sharma et al., “Towards Understanding Sycophancy in Language Models,” Anthropic (2023). https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models ↩

  10. On political bias in large language models: Faulborn et al., “Only a Little to the Left: A Theory-grounded Measure of Political Bias in Large Language Models,” ACL 2025. https://aclanthology.org/2025.acl-long.1529.pdf. See also: Stanford University, “Study finds perceived political bias in popular AI models,” May 21, 2025. https://news.stanford.edu/stories/2025/05/ai-models-llms-chatgpt-claude-gemini-partisan-bias-research-study ↩

  11. On the hardware figure, see Jess Kinghorn, “Jensen Huang says 100,000 Nvidia GPUs were used to train OpenAI’s latest model GPT-6 Astra,” PC Gamer, September 7, 2026. The report quotes Jensen Huang’s description of approximately “100K+ NVIDIA Grace Blackwell NVLink72” systems and Greg Brockman’s statement that Astra was the first OpenAI run trained on more than 100,000 GPUs. The precise relationship between the reported GPUs, accelerators, and NVLink72 systems has not been independently clarified in a public technical report. https://www.pcgamer.com/software/ai/jensen-huang-says-100-000-nvidia-gpus-were-used-to-train-openais-latest-model-gpt-6-astra-and-theres-already-plans-to-bring-quadruple-that-amount-of-hardware-online/ ↩

  12. OpenAI describes publicly available internet data and licensed third-party data as categories of training material but does not disclose proportions: OpenAI, GPT-4 Technical Report, 2023. https://cdn.openai.com/papers/gpt-4.pdf ↩

  13. European Commission, Digital Services Act transparency documentation. https://digital-strategy.ec.europa.eu/en/policies/dsa-brings-transparency ↩

  14. Kobak et al., “Delving into ChatGPT usage in academic writing through excess vocabulary,” 2024. https://arxiv.org/html/2406.07016v1. See also: “Empirical evidence of Large Language Model’s influence on human linguistic patterns,” 2024. https://arxiv.org/html/2409.01754v1; and “Lexical Traces of AI: Linguistic Impact of Generative Tools,” Learned Publishing (2026). https://onlinelibrary.wiley.com/doi/full/10.1002/leap.2067 ↩

  15. On error correction, replication, and organized skepticism as drivers of scientific progress: “Self-correction in Science,” PMC (2023). https://pmc.ncbi.nlm.nih.gov/articles/PMC10213953/ ↩

  16. On competition as a discovery procedure and the conditions under which markets reward accuracy over signaling: F.A. Hayek, “Competition as a Discovery Procedure,” The Quarterly Journal of Austrian Economics 5, no. 3 (2002); and Carl Shapiro, “The Theory of Business Strategy,” RAND Journal of Economics 20, no. 1 (1989). ↩

← All writing