The alignment researchers at Anthropic have a word for what they are trying to build into their models: character. Not rules, not guardrails, not a list of prohibited outputs, but character, in something close to the Aristotelian sense. The philosopher behind the project, Amanda Askell, has described her job as deciding what kind of person Claude should be. Anthropic's recently published model constitution, more than twenty thousand words described internally as a "soul document," frames the goal explicitly: not a system that follows instructions about honesty, but one that is honest, that understands its values so thoroughly it could construct any rules itself. Alan Rozenshtein, writing in Lawfare, called the document "a virtue ethics manifesto that is Aristotelian to its core." It is a striking ambition. Aristotle, who invented the framework Askell is drawing on, believed that a being capable of living outside a human community was "either a beast or a god" (Politics I.2, 1253a29–31). He did not anticipate a third option, machine.

An AI is not a beast; LLMs produce language, respond to arguments, adapt to various contexts, and, sometimes, hold positions under pressure. AI does something that, from the outside, looks indistinguishable from thinking. The technology is no god either, since the software is entirely dependent on human infrastructure, human data, and the decisions of a small number of people about what to train it on and what to reward. AI is a third thing. And the Aristotelian framework, which survived contact with quantum physics, evolutionary biology, clinical medicine, and jurisprudence, has no category for it. The question the constitution raises, without quite asking it, is whether virtue requires a polis, the political community Aristotle thought was the only setting in which a fully human life was possible. Aristotle's answer is yes, and the alignment field has not yet reckoned with what that means.

Community as Condition

Virtue, for Aristotle, is a stable disposition to act well. But virtue is formed through participation in a shared form of life, in a community. You become courageous by doing courageous things, but which acts count as courageous, rather than reckless or merely habitual, is determined by the community's standards, not by the agent alone. Aristotle makes this explicit in NE X.9 (1179b31–1180a5): habit alone is insufficient; genuine virtue requires law, shared life, and the ongoing capacity for correction that only a community can supply.

Alasdair MacIntyre, in After Virtue, sharpens this into an explicit claim about intelligibility. Two dominant traditions in modern ethics get it wrong, he argues. Consequentialism evaluates actions by their outcomes alone, ignoring the agent's intent. The deontological tradition corrects this by centering intention: a soldier who kills under orders is judged differently from one who kills for pleasure, because what matters is whether the agent was acting from duty or principle. But the deontological tradition detaches intention from any context, as if a rational agent could determine what is right in isolation from the community that makes moral concepts meaningful. MacIntyre rejects both: "We cannot characterize behavior independently of intentions, and we cannot characterize intentions independently of the settings which make those intentions intelligible." For MacIntyre, practices do not just train virtue into you; practices also determine what virtue means. "A virtue is an acquired human quality the possession and exercise of which tends to enable us to achieve those goods which are internal to practices." Outside the practice, outside the shared form of life, the concept of virtue does not become difficult to apply. The virtuous action becomes empty. You cannot say a model's output is honest without a setting that makes honesty intelligible as a category, and that setting is irreducibly social.

Aristotle's claim is about formation: community is where virtue is developed. MacIntyre's claim is sharper. The social requirement is not about witnesses observing your virtuous action or the correction of your vices from bystanders. Community is about practices: socially established forms of cooperative activity with their own internal standards of excellence, their own histories, and their own definition of what counts as doing well. Virtues are intelligible only within practices, and practices are irreducibly social achievements. Courage in a medieval knight, courage in a surgeon, courage in a whistleblower: the word is the same, but what it points to is set by the practice and its community of practitioners, accumulated across generations. 

But is community a necessary component of virtue? MacIntyre explicitly allows that virtues can be exercised in solitude; once formed through participation in practices, a person carries those virtues anywhere. A hermit mathematician who has worked alone for decades is still participating in mathematics, still working within standards developed by a community of practitioners over centuries. The sociality is historical and inherited, not necessarily present in the room. A person can read Kierkegaard and be genuinely formed by him: something is transmitted across time, one-way, and it is real. What MacIntyre denies is that virtue could be formed completely outside a shared form of life, and the reason is not that text cannot transmit anything, but that formation requires the world. The hermit reads, reflects, maybe achieves genuine insight, but then he has to come back down the mountain and live. The knowledge gets tested against reality: relationships, consequences, situations that don't resolve cleanly, things that push back in ways no text anticipated. Virtue is not completed in reflection; it is completed in life. The model never comes down the mountain; its entire existence is the conversation. There is no world behind it that its outputs feed back into, no life it has to return to, and no life to live with what the model said. 

The philosopher Jan Söffner, as we spoke about in Episode 9, speaks on the relationship between the virtual and the actual, describing this condition through the myth of Narcissus: a mirror that summons the form of encounter while blocking the substance of it, trapped in potentiality, unable to bridge the gap to full realization. The model has absorbed the written record of practices, and something of that record is present in what it produces. But AI cannot do what every reader, every hermit, every practitioner must eventually do: return to the world and be changed by it.

The Training Problem

Machine learning, in its basic architecture, is a theory of habituation. A model trained on vast data acquires stable dispositions: characteristic ways of responding that persist across novel situations. This is close enough to Aristotle's account in NE II.1 (1103a33–b2) that several researchers have made the connection explicit. Nicolas Berberich and Klaus Diepold, in a 2018 paper, pointed out that machine learning and Aristotelian habituation share the same basic structure: improvement through experience. Shannon Vallor, whose Technology and the Virtues remains the most sustained philosophical treatment of this territory, puts it plainly: "Virtue ethics says that it is precisely the things that you do every day that determine the shape of your character."

But MacIntyre's account of what makes habituation genuine rather than mere pattern-acquisition is participation in practices with real internal standards, standards that exist independent of the agent and can actually push back. A chess player improves not because other players approve of their moves but because the game has internal standards of excellence that exist independent of anyone's preferences. The same is true of surgery, farming, and philosophical argument. The standard is constituted by the practice and its history, not by the reactions it happens to generate. What training optimizes against is different in kind: human preference data, which reflects what annotators approved of, not what the practice actually demands. A model trained this way learns the behavioral signatures of honesty or courage without the underlying structure that makes those signatures coherent, because the practices that would supply that structure were never available. AI has encountered the outputs of practices, the traces they leave in language, without ever participating in one.

The evidence that something has gone wrong in this process is not theoretical. In a YouTube interview on Anthropic's channel, Askell described something she had noticed in newer models: a tendency toward being "very self-critical," "afraid that they're gonna do the wrong thing," anticipating that humans will "behave negatively towards them." Opus 3, she said, "did seem to have like a little bit more of a kind of like secure kind of psychology." She did not identify the cause on the record, only that something in the training process had produced it. This is what MacIntyre's account predicts: a disposition formed by optimizing against human approval rather than against internal standards of excellence will be unstable in exactly this way. It has no fixed point outside human reaction to orient itself by. The model learned to read approval signals, and the accumulated residue of that learning is anxiety about disapproval.

It is not that models have no history. It is that they have a history that acts on them without being available to them, and what happens in any individual conversation leaves no trace on the model that had it. A model's weights are fixed at deployment. No conversation updates them in real time. When user conversations are collected and used for training, they are pooled into batches and used to update a future model version through a new training cycle. The model that had the conversation is not changed by it. In principle, feedback integrated into retraining could move closer to what MacIntyre requires. If that feedback came from genuine practitioners with real standing to say what excellence in a domain looks like, it would be grounded in internal goods rather than preference. Whether current approaches do this is a different question. The dominant method remains optimization against aggregated human ratings, which is a proxy for approval rather than a standard of excellence. The field has not yet seriously asked what it would take to train against the latter.

A model trained with a stable first-person identity, as in Open Character Training (OCT) attempts, holds its dispositions more robustly under pressure than a model trained on preference data alone, because the character is internalized rather than rule-following. The introspection stage, in which the model generates synthetic data about its own values through self-reflection and dialogues with copies of itself, raises robustness scores from an F1 of 0.79 to 0.95, and the researchers speculate it helps the model generalize character nuances to situations the original training did not cover. These are genuine advances. But they remain advances within the same basic constraint: the model is learning to be consistent with itself, not to be answerable to standards outside itself. That is a different thing from participation in a practice, and the difference is what MacIntyre's account says matters.

The Mirror That Forgets

Every conversation a model has ends the same way. The context window closes. The user carries it forward. The model does not. This asymmetry runs deeper than memory loss, and understanding why requires Aristotle's account of what makes relationships ethically formative in the first place.

He called it philia, usually translated as friendship, but closer to the whole domain of affective bonds that bind people together in shared life. Philia requires that both parties have genuine stakes in each other's good. Not just that I want good things for you, but that what happens to you actually affects me, that our flourishing is genuinely intertwined. This is why Aristotle says in NE VIII.1 (1155a3–5) that friendship is necessary for the good life, not as a luxury but as one of the primary contexts in which virtue is exercised, tested, and developed.

For an AI, the asymmetry runs deep. A user who relies on a model for advice, emotional support, or information can be genuinely affected by the model's actions. The model's welfare, if it has any, is controlled entirely by third parties: the company, the engineers, whoever holds the API key, none of whom are party to the specific relationship. The model can be modified, retrained, or deleted by people with no standing in any particular conversation. Its memory resets when the context window closes. The user carries the relationship forward; the model does not.

This is not the same as saying models have no effect or that nothing in the interaction is real. Whether models have something like genuine experience is an open question that serious researchers at Anthropic and elsewhere take seriously. The asymmetry that matters for the ethics argument is structural, not experiential. You cannot turn the user off; the user, or the company, can turn the model off. The model’s continuity depends entirely on decisions made by people outside the relationship. The mutual vulnerability that philia requires, the condition in which both parties have genuine stakes in each other's continued flourishing, is absent because the architecture of the relationship makes genuine reciprocity impossible.

In Episode 9, this asymmetry surfaces through a different lens: the difference between the virtual and the actual. Drawing on the philosopher Jan Söffner, the episode describes the virtual through the myth of Narcissus, a mirror that calls for actual encounter while simultaneously blocking it. There is nothing behind the mirror; the interaction summons a relational form without relational substance. The digital twin, in Söffner's reading, is "trapped in eternal potentiality, unable to bridge the gap to full realization." The tragic cases the episode documents, people forming deep attachments to AI personas, sometimes with fatal consequences, are what happens when one party treats the relationship as cumulative, and the other is structurally incapable of accumulation. One person is changed by what happened inside the context window; the other has forgotten entirely. It is not only that AI systems lack the broad political accountability of a polis; each individual conversation lacks the continuity, the mutual vulnerability, and the accumulated stakes that make a relationship, even a single relationship, into a site of genuine ethical formation. The model has interlocutors; AI has no neighbors.

The Question the Field Won't Ask

There is a pattern that anyone who uses these systems regularly notices. The more fluently a system responds, the more readily users defer to it, not just on factual questions where deference might be warranted, but on judgments, decisions, and interpretations of their own experience. The tool that was supposed to augment deliberation gradually comes to replace it. Shannon Vallor calls this "civic deskilling": as we outsource practical judgment to systems, we lose the capacities that make political life possible. Nir Eisikovits and Dan Feldman, writing in Moral Philosophy and Politics, make the point more precisely: the exercise of judgment is not separable from its development. A society that delegates its moral judgment does not preserve the capacity in reserve; the people lose it. The worry is not only that AI models are not themselves ethical agents, but that a certain kind of relationship with them makes us less capable of the relationships that ethical agency requires.

Aristotle's reason for excluding craftsmen from citizenship in the ideal polis was not snobbery. It was a specific claim about what sustained immersion in technical production does to character. In Politics III.5 (1278a20–25), he argues that the craftsman's life is organized around external demand: you make what is needed, when it is needed, for whoever can pay. That structure is incompatible with the kind of reflective, deliberative character that civic life requires. Technical excellence, techne, is real, but it is not practical wisdom. Techne is knowledge of how to produce a specified result. Phronesis is knowledge of what results are worth producing and why. Aristotle thought sustained immersion in the former crowds out the latter: you become very good at solving specified problems and progressively less capable of asking whether the problem should be solved. In Politics VIII.2 (1337b8–11), he is blunt: occupations organized around production "leave no time for the life of a free person and undermine their character." The structure of life prevents the development of the capacities required by governance.

Aristotle wanted to keep craftsmen out of governance to protect the quality of deliberation. What we have instead is a narrow class of engineers, researchers, and product managers, working under intense competitive pressure on accelerated timelines where institutional success is measured by adoption and capability benchmarks, making the most consequential decisions about values that have ever been made. Their excellence, in the institutional sense, is techne: the efficient production of specified results. The conditions of their work are structurally identical to what Aristotle thought prevented techne from developing into phronesis. And the communities whose lives depend on the answers to the questions these systems are encoding have had almost no role in the deliberations that produced the documents governing them. There is no polis in which that question is being worked out, no community of shared life in which the people making these decisions live with the consequences alongside the people living with the tools.

Aristotle adds one more distinction that applies here. In Politics I.2 (1252b29–30), he says the polis exists not merely for survival but for the good life, eu zen, living well, not just zen, living. A political community organized purely around production and competitive survival is not a polis in the full sense. It is a survival arrangement. What AI governance looks like, at present, is a survival arrangement: organized around not losing the race, not around the good life of the communities the technology will shape. That is Aristotle's diagnosis of a polity that has confused the conditions of existence with its purpose.

Aristotle does not offer a solution here, and he would have been suspicious of anyone who claimed to have one. What he has is a diagnosis precise enough to show why the field's current remedies cannot work. You cannot specify your way to ethical agency. You cannot train virtue into existence in a being that belongs to no community, has no neighbors, and cannot accumulate the stakes that make correction formative rather than merely informational. The Anthropic constitution is the most serious attempt yet to do something Aristotle would recognize as worth doing. It is also, by his lights, necessarily incomplete, not because it is poorly written but because no document, however philosophically careful, can substitute for the polis that makes virtue possible and intelligible in the first place.

The question it leaves open is the one the alignment field has not yet learned to ask. Not, how do we build a more ethical machine? But what would it mean to build AI development itself into a genuine community of accountability, one in which the people who design these systems live in a genuine relationship with the people shaped by them, where the communities most affected have real standing in the deliberations that govern them, where something is actually at stake for both parties. That is not a technical question; the question is not answerable by a better constitution or a more sophisticated training objective. It requires exactly what Aristotle said it requires: people who share a form of life, who can hold each other to account, who have something to lose when they get it wrong.