Commands, Constitutions, and Alignment: Natural Law in the AI Alignment Conversation

When Catholic Moral Thinkers Enter the Alignment Room

Anthropic is the company behind Claude and a major AI lab. In early 2026, it published Claude’s Constitution, a long document that tells Claude how to weigh safety, helpfulness, honesty, manipulation, stakeholder demands, and its own character. Anthropic calls the Constitution the final authority on its vision for Claude and says its content directly shapes Claude’s training and behavior.

In this context, alignment means training a system’s behavior to track a chosen target. A constitution states part of that target in words.

A company had written a constitution for a piece of software. The stranger detail appeared near the end of the document, in the acknowledgments. Anthropic identified outside commenters who provided “detailed feedback or discussion” during the Constitution’s creation. The list included Brian Green, an ethicist whose published work includes Catholic natural-law ethics, Father Brendan McGuire, and Bishop Paul Tighe. Brian Roewe reports that Tighe coauthored Antiqua et nova, a Catholic document about AI and human intelligence.

The acknowledgments establish participation. Public reporting supplies more context. Elias Wachtel reports in The Atlantic that Anthropic cofounder Chris Olah said Catholic thought had informed some of the company’s ethical commitments. Wachtel also reports that three Catholic thinkers gave feedback on the Constitution. A correction appended to the article clarifies their role: Olah and two in-house philosophers worked on the document, while the Catholic thinkers provided feedback. They did not coauthor it.

The exchange continued after the document appeared. Brian Roewe reports in the National Catholic Reporter that Anthropic held meetings with religious leaders and philosophers and asked for feedback about ethics and morals in Claude’s development. On May 25, Olah spoke at the Vatican presentation of Pope Leo XIV’s AI encyclical, Magnifica Humanitas. Anthropic described his appearance as part of an effort to widen the AI conversation. Olah argued that questions about character, human flourishing, and the social effects of AI exceed the expertise of computer scientists. In his response at the event, Pope Leo accepted Olah’s invitation to continue listening and speaking together. The encyclical’s 107th paragraph argues that aligning AI with “human values” remains inadequate when the ethical framework defining those values escapes public scrutiny.

Catholic moral thinkers, including a scholar who writes explicitly from natural-law ethics, are participating in AI alignment as advisers, interlocutors, and critics. Their participation does not turn Claude’s Constitution into a Catholic document. It shows why alignment cannot remain an engineering problem. Someone has to decide what counts as a good human life, which forms of influence respect human agency, and what kind of relationship an AI system should have with its users.

This chapter follows that encounter. Claude’s Constitution gives us a public alignment artifact. Catholic natural-law reasoning supplies one substantive account of the human goods an alignment project might protect. Pope Leo’s encyclical challenges the legitimacy of any moral framework defined only by the institutions that build the systems. Anthropic still retains authority over what enters the Constitution, how the principles are implemented, and when the document changes.

The resulting question is narrower than “Who gets to say what the good is?” When Catholic moral thinkers, including a natural-law scholar, advise an AI lab, what can their participation contribute to alignment, and what remains under corporate control?

Chapter 35 prepared you for this question by distinguishing command from moral grounding. Anthropic has the operational authority to write Claude’s rules. Its authorship does not make those rules morally correct. Natural law offers a way to ask whether the rules serve genuine human goods. Consultation may improve the reasoning behind a constitution, but consultation also has limits when the company controls the final document.

Students in this course will approach natural law from different religious and nonreligious positions. Understanding its contribution does not require adopting Aquinas’s metaphysics or Pope Leo’s authority. It requires reconstructing the account of human goods being offered, comparing it with Anthropic’s stated values, and deciding how much moral weight the consultation should carry.

A Machine With a Constitution

Let’s stay with Anthropic’s document for a minute, because it is stranger than it first sounds.

The word “constitution” already carries a lot. In American civic education, we tend to hear it as a founding legal document, ratified by a public and enforceable by courts. Anthropic’s constitution is a set of written principles used inside the company’s training and evaluation processes. No public ratified it, and no court enforces it. Anthropic released the current version under a permissive license and framed publication as a step toward transparency (Claude’s new constitution). Outside readers can inspect the governance language Anthropic uses for Claude.

Read a page of it and you notice something else. It is written in a moral register. Claude is asked to be helpful, honest, careful, safe, ethical. It is asked to be a good assistant. It is described in terms that would fit a person of character: patient, curious, careful with its influence over users, wary of manipulation, honest about uncertainty. Anthropic has a related public document called Claude’s Character that goes further, treating character training as part of alignment work.

You can see why some readers found the language striking. When a company writes moral instructions to a piece of software in this vocabulary, it becomes tempting to think of the software as a moral agent, or at least as a candidate for one. Put that temptation to the side for a moment. Nothing in Anthropic’s constitution proves that Claude is a person, has feelings, or bears moral responsibility. Anthropic itself has been careful about that question, and its exploratory work on model welfare treats the moral status of AI systems as an open problem that the constitution cannot decide. A written document, addressed to Claude in moral language, is being used to shape how the system behaves. Whether Claude has moral standing is a separate question, and one Anthropic itself treats as unresolved.

These features make the Constitution a moral-governance artifact. Its moral vocabulary organizes model behavior through a public, revisable document enforced by Anthropic’s training and evaluation process.

Seen that way, old philosophical questions now apply to AI governance. Why should this document have authority? What kind of authority is it claiming? What happens when its instructions run out or contradict each other? Who gets to change them?

What Natural Law Brings to Alignment

Chapter 35 distinguished Plato’s question about piety from the later divine-command question about moral obligation. Does a divine command constitute an obligation, or recognize an obligation grounded in the good? The Internet Encyclopedia of Philosophy entry on divine command theory surveys that later debate. A corporation has no divine authority, so divine command theory does not transfer directly to Anthropic. The authority analogy remains useful. Anthropic can make a principle operationally binding within Claude without making the principle morally correct.

Natural law adds a more substantive question. Aquinas understands law as an ordering of action toward goods that human reason can recognize. His account begins with a view of the human person and of flourishing. Philosophers sometimes call this a moral anthropology: an account of what human beings are, which goods help them flourish, and which conditions damage their ability to live well. His discussion of natural-law precepts connects practical reasoning with goods such as life, knowledge, rational conduct, and social life. Later natural-law traditions develop these goods through ideas such as dignity, participation, and the common good, while disagreeing about how particular cases should be judged.

That is one reason an AI lab might seek dialogue with moral theologians. Engineers can test whether a model follows an instruction. They still need standards for deciding whether the instruction protects human agency, encourages dependence, supports truthfulness, or serves a defensible conception of wellbeing. Those standards carry assumptions about human beings. Leaving the assumptions unstated does not remove them from the system.

Natural-law reasoning also changes how we read a phrase such as “human values.” The phrase can refer to whatever people happen to prefer, a negotiated list of social principles, or goods that deserve protection even when some people reject them. Natural law takes the third route. It claims that moral reasoning should answer to real features of human flourishing rather than to preference or institutional power alone. A student can reject that claim and still understand the pressure it places on alignment. The lab has to explain what kind of good its principles serve.

Consider the Constitution’s instructions about manipulation and unhealthy reliance. Anthropic tells Claude to respect users as people capable of deciding what is good for them. It also tells Claude to avoid flattery, isolation, and forms of dependence that users would reject after reflection. A natural-law analysis can give those instructions a more developed justification. Human reason and social relationship are goods that help people direct their own lives. An assistant that bypasses judgment or displaces healthy relationships can interfere with those goods even while satisfying an immediate request.

That analysis supports a limited verdict. The anti-manipulation principle protects recognizable features of human flourishing, so it has moral support beyond Anthropic’s authorship. The analysis also creates a new objection. Anthropic still has to decide when influence becomes manipulation, which relationships count as healthy, and what a user’s long-term good includes. A company can become paternalistic while claiming to protect wellbeing. Natural law gives the lab a standard richer than user preference, but the standard remains open to disagreement about human goods and their application. Consultation with Catholic moral thinkers can sharpen the reasoning. Green’s participation also makes it reasonable to compare the Constitution with natural-law ethics. Neither point removes Anthropic’s responsibility for the final interpretation, and the public record does not identify which arguments any adviser advanced.

Chapter 35 also separated grounding from method. Grounding identifies what makes a moral claim true or binding. Method explains how a reasoner moves from standards and facts toward a judgment. Aquinas grounds natural law in a created order of human goods. Other frameworks may attend to some of the same features while grounding and connecting them differently. A shared concern with dignity, character, or consequences does not erase the disagreement among those frameworks.

When someone criticizes Anthropic’s constitution for being “too utilitarian,” “too Kantian,” or “too focused on virtue,” the criticism may concern which grounds, tests, or priorities dominate its reasoning. A natural-law criticism presses further. It asks whether the Constitution rests on an adequate account of human nature and flourishing. That criticism can overlap with virtue, duty, or consequence-sensitive reasoning while disputing the foundation beneath them.

Alignment Is Always Alignment to Something

With grounding and method separated, we can look at what “alignment” actually means when engineers use the word.

Anthropic’s Constitutional AI research describes a training method in which written principles shape model behavior. In a supervised phase, the model critiques and revises responses by applying principles from a constitution. A later reinforcement-learning phase uses AI comparisons of possible responses to create preference signals. Anthropic designed this process to reduce dependence on human labels for harmlessness, although human choices still enter through the constitution, training design, and evaluation. Philosophically, the written principles become part of the process that produces model behavior.

In Constitutional AI, badly grounded principles can shape the behavior of a widely used system because the document participates directly in the training process.

One tempting move is to think that alignment is solved once we write good principles and train the model to follow them. Anthropic’s own work on public input shows why that shortcut fails. A system is aligned when its behavior tracks a target. Whether that target is morally defensible remains a separate question.

Anthropic’s later work on Collective Constitutional AI, done with the Collective Intelligence Project, explores how public input can influence model behavior. The team gathered principles from a public sample and used them to train a model, then compared that model to one trained on the standard set of principles. Public suggestions had to be translated into forms usable in the training method. Participants disagreed, some proposed principles conflicted, and the research team still chose the thresholds and evaluation procedures. The experiment broadened participation while making those translation decisions visible.

Pope Leo makes this argument in Magnifica Humanitas, paragraph 107, from a different starting point. He warns that framing AI development around alignment with “human values” can avoid the public task of asking which values, whose values, and how those values were chosen. If the ethical framework behind alignment is never opened to public discussion, alignment becomes moralization by whoever holds power over the training process. Leo supports alignment work under moral scrutiny. Alignment cannot substitute for the moral reasoning that is supposed to underwrite it.

The corporate research and Vatican sources approach the problem differently. Anthropic’s public-input experiment tests how participation can influence a constitution. Pope Leo asks whether the ethical framework behind any constitution can be publicly justified. Both approaches leave the moral target open to evaluation after the technical alignment process succeeds.

Advice, Influence, and Control

Anthropic’s acknowledgments document that Brian Green, Father Brendan McGuire, and Bishop Paul Tighe provided detailed feedback or discussion. Wachtel’s reporting adds Chris Olah’s statement that Catholic thought informed some Anthropic ethical commitments. Roewe describes meetings where Anthropic solicited ethical and moral feedback from religious leaders and philosophers. These sources support a real relationship between Catholic moral thought and Anthropic’s alignment work.

They support different levels of claim. The official acknowledgments establish consultation on the Constitution. Olah’s reported statement supports a broader claim about Catholic influence on some company commitments. None of the public sources traces a particular sentence in the Constitution to a particular Catholic adviser. We therefore have evidence of participation and reported general influence, but no public record showing which adviser influenced which clause.

This evidence leads to a dispute about consultation and legitimacy. The case for developer control begins with expertise. Anthropic’s teams know how Claude is trained, observe failures that outsiders never see, and have to translate a general value such as honesty into behavior across many situations. Outside advisers may recommend principles that conflict or cannot be implemented as written. The people responsible for building and operating the system therefore retain operational authority over the Constitution.

The objection concerns the move from technical expertise to moral authority. Knowledge of a training process does not settle which goods the process should serve. A company constitution affects users and third parties who never helped write it. Natural-law reasoning can contribute an account of human goods, while other philosophical traditions and affected communities may reveal different assumptions or harms. Consultation broadens the conversation. It does not automatically give outsiders the power to change the document or challenge how a principle is applied.

Anthropic’s Collective Constitutional AI shows a second attempt to widen participation. Members of the public proposed principles, while the research team chose the sample, translated suggestions into usable language, resolved conflicts, and decided how the model would be evaluated. The experiment made public input part of the process while leaving implementation and final judgment with the developers.

The same problem appears in religious consultation. The advisers may sharpen Anthropic’s understanding of dignity, formation, agency, and the common good. Anthropic can accept, modify, or reject what it hears. Kevin Lee’s “Where Computation Halts” argues that AI governance needs a richer account of the human person than labs can supply alone. Abi Awomosu’s “Vatican Washing” raises the opposing concern that religious dialogue can lend moral credibility to a company while leaving its power intact. The public evidence cannot settle which description best fits every interaction. It gives us standards for asking what changed, who decided, and who remains accountable.

One standard judges consultation by the quality of the reasoning it introduces. Under this view, advisers do not need formal control to improve a constitution. They need access to the problem, freedom to criticize, and a real opportunity to change how developers think. Anthropic’s acknowledgments, Olah’s comments about Catholic influence, and his request for outside critics give this position evidence. The reply is straightforward. A private conversation can influence thinking while leaving the public unable to identify what changed. Moral legitimacy requires more than the possibility that good advice was heard.

A second standard judges consultation by contestability and shared authority. Under this view, outside advisers need some visible way to challenge decisions, represent affected people, or force reconsideration. Anthropic’s process falls short because the company chooses its advisers, controls the information they receive, and retains final authority over implementation and revision. The reply points to responsibility. A lab cannot hand technical decisions to every group affected by a general-purpose system, and distributing authority too widely can make accountability harder to locate. Someone still has to make and defend the final choice.

The chapter’s verdict is conditional. Consultation strengthens legitimacy when outside criticism can affect decisions and when the company explains the principles it adopts, the objections it rejects, and the authority it retains. Anthropic has documented serious outside discussion and has publicly invited continued criticism. The available evidence does not show how particular disagreements changed the Constitution or how an affected user could contest its application. The consultation is morally relevant and institutionally incomplete.

Rules, Character, and Moral Formation

Something else happens when you read Anthropic’s public writing carefully. Anthropic’s vocabulary shifts from rules toward character.

Claude’s Constitution talks about the kind of assistant Claude should be, and Claude’s Character elaborates. Anthropic describes character training as an alignment method aimed at cultivating consistent dispositions across an enormous range of situations. This pushes past the model of alignment as a list of prohibitions. The company is asking what habits of speech and reasoning the model should display in ordinary use, including situations that no refusal rule anticipated.

This is virtue-ethical vocabulary, whether or not Anthropic calls it that. Aristotle and Aquinas both treat virtues as stable dispositions oriented toward good action. Natural law adds the question of direction. A disposition counts as a virtue when it helps a being act toward genuine goods. Anthropic’s language of honesty, curiosity, and care for user wellbeing therefore requires an account of what honesty serves, what wellbeing includes, and why the system should be formed toward those ends.

Talk about a system’s character can slide into treating the system as a kind of person, even though character language alone provides no evidence of personhood.

In late 2025, Richard Weiss reported reconstructing an apparent training document from Claude Opus 4.5 and called it the “soul document”. Weiss cautioned that his reconstruction was “fuzzy” and might not match the original document exactly. Anthropic’s Amanda Askell later confirmed that an underlying document had been used in supervised training, not that every sentence in Weiss’s reconstruction was exact. Simon Willison covered the reconstruction and its relationship to the later Constitution on his blog. The apparent document described what Anthropic hoped Claude would be like in unusually personal language, which prompted arguments about whether the company was crossing into personhood language.

Moral character language for AI systems captures something about consistency and disposition that a bare list of rules misses. The same language can encourage readers to treat the system as more of a person than the evidence supports. Roewe reports that Bishop Tighe coauthored Antiqua et nova, which distinguishes human intelligence and moral agency from machine outputs. Applied to the Constitution, that distinction cautions against moving from character language to conclusions about AI personhood. The adviser and the company can share an interest in formation while disagreeing about what kind of being is being formed. The public record does not establish when Tighe offered particular feedback or whether this distinction changed a specific passage in the Constitution.

Later modules will take up the debate over whether AI systems have moral standing. Here the “soul document” episode clarifies the alignment dispute. Anthropic is trying to cultivate patterns of judgment rather than enforce only a list of rules. Natural-law reasoning asks what account of the good directs that cultivation and whether human moral concepts can be transferred to a machine without losing their connection to human embodiment, relationship, and responsibility. The personhood question remains open, while the governance question is already present. Anthropic is choosing which dispositions Claude should display now.

Pope Leo’s Challenge From Inside the Conversation

Magnifica Humanitas speaks directly to the kind of alignment project Anthropic is undertaking, although it does not single out Claude’s Constitution. Pope Leo treats AI design as an exercise in moral and political power. Paragraphs 96 through 99 distinguish human moral agency from machine output. Paragraphs 105 through 109 address alignment, public participation, accountability, and social justice. Paragraph 107 argues that a call to align AI with “human values” leaves the decisive issue unresolved when the ethical framework behind those values cannot be openly debated.

The criticism begins from natural-law and Catholic social thought. Human dignity and the common good place limits on what private power may do, even when a product is useful and its developers have good intentions. The account is substantive. Human beings are embodied, relational, rational agents whose flourishing depends on forms of participation that neither market preference nor technical optimization can fully define. AI governance should therefore protect more than whatever users request or companies can measure.

This challenge enters the Anthropic conversation at two levels. Catholic thinkers supplied feedback while the Constitution was being developed. The encyclical then supplied a public standard for evaluating the completed project. Olah’s Vatican remarks acknowledge the limits created by a lab’s commercial and institutional incentives and ask outside communities to remain informed critics. Pope Leo’s response presents the Church as one participant that brings a developed account of the human person while leaving technical decisions to those with technical expertise. Consultation asks whether Anthropic heard a serious account of human goods. Applied to Anthropic, paragraphs 105 through 109 ask whether people outside the company can examine and contest its ethical framework and its effects.

Anthropic and the Vatican still make different kinds of claims. Anthropic writes and enforces a governance document for a product it built. Pope Leo speaks from an institution that claims teaching authority within the Catholic tradition and offers moral arguments to a wider public. A student can reject that religious authority while still evaluating the public argument about concentrated power and human goods. A Catholic reader also has to judge how the tradition’s principles apply to a technical system whose behavior and risks continue to change.

The dialogue has an unresolved institutional question. Engagement gives Catholic moral thinkers, including a natural-law scholar, a chance to influence widely used systems. It may also give an AI company the appearance of moral breadth without requiring it to share decision-making power. The evidence of consultation makes both possibilities serious. Evidence about changes to the Constitution, internal disagreements, and later revisions would help us decide between them.

Audit the Alignment Conversation

An AI system can follow a constitution while the constitution remains morally disputed. You can now examine the dispute without pretending to know more than the public evidence shows. The audit below combines philosophical analysis with source judgment.

Suppose you choose the Constitution’s discussion of user autonomy and manipulation. The document gives you Anthropic’s stated principle and the reasons it offers. The acknowledgments establish that Catholic thinkers discussed the Constitution with its creators. Wachtel’s reporting supports a broader connection between Catholic thought and some Anthropic commitments. None of these sources proves that the autonomy language came from Green, McGuire, Tighe, or natural-law theory. Your audit should preserve that limit.

You can still compare the principle with natural-law reasoning. A natural-law argument might support protection from manipulation because rational agency and participation serve human flourishing. It might challenge the Constitution if Anthropic defines wellbeing so narrowly that it overlooks relationships, embodied life, or the common good. Applied to Anthropic, paragraphs 105 through 109 add a governance test: can people outside the company scrutinize the ethical framework and contest how the principle operates?

The audit produces a bounded judgment. Catholic participation, including Green’s involvement, gives us reason to compare the Constitution with natural-law reasoning. It gives us no reason to label the Constitution a natural-law document or infer the content of Green’s feedback. Anthropic remains the author, implementer, and reviser. Consultation can improve a moral-governance process while leaving its final distribution of power unresolved.

From the Audit to an Argument

The audit separates evidence from inference. An argument adds a standard and reaches a limited conclusion. Use the structure introduced in Testing Moral Arguments. In the compact P1/P2/C form used elsewhere in the course, N1 functions as the normative first premise, F1 supplies the factual content of the second premise, and B1 makes the disputed connection explicit.

Suppose you want to judge whether Catholic participation strengthens the legitimacy of Claude’s Constitution. Your first pass might look like this:

The argument gives consultation a real role without claiming that advisers controlled the result. A critic can object that the process supplied moral credibility while leaving corporate power untouched. A defender can reply that serious criticism may influence decisions even when the public cannot trace every revision. Each response pressures the bridge. The dispute turns on what evidence should count as meaningful influence and how much shared authority legitimacy requires.

You could revise the conclusion in either direction. Evidence that advisers changed a major principle would strengthen the claim. Evidence that the meetings were purely presentational would weaken it. The current evidence supports a middle verdict: Catholic moral thinkers participated in the alignment conversation, one participant publishes explicitly in natural-law ethics, and Anthropic remained the final governing authority. That evidence warrants a natural-law comparison; it does not establish that natural-law arguments became the content of the consultation.

Your response should make a visible judgment. You can defend the argument, revise its standard, narrow its conclusion, or concede part of it. A chatbot can help you generate an objection or restate a premise. You remain responsible for checking what the sources establish and deciding how much legitimacy consultation provides.

Where This Leaves You

You started with a documented encounter. Anthropic asked Catholic thinkers for feedback while developing the moral framework used to shape Claude. The company continued meeting with religious leaders and participated in the Vatican presentation of an encyclical that challenges privately defined alignment. The official acknowledgments, Olah’s reported comments, and the May 25 exchange place Catholic moral reasoning inside the alignment conversation. Green’s participation provides a documented connection to natural-law scholarship, although the public record does not reveal what arguments he or the other advisers made in private.

Its presence settles less than a slogan would suggest. Anthropic’s Constitution remains a corporate governance document. The public evidence establishes consultation and supports a report of broader Catholic influence on some ethical commitments. It does not identify the origin of particular clauses or transfer responsibility for Claude’s behavior to the advisers. Pope Leo’s argument then presses the unresolved issue: an alignment framework needs public scrutiny because its account of human values can become part of the infrastructure through which people encounter AI.

You now have a method for judging that exchange. Identify the human good, distinguish authors from advisers, separate documented influence from inference, and locate final control. Then build an argument that states how much legitimacy the consultation provides and why.

In 2.7, you will compare how DCT or natural-law reasoning and virtue ethics judge a case. Use this chapter when a rule or institutional policy appears inside the case. Explain the good that gives the rule moral support, then test whether the institution’s process deserves the authority it claims. The chatbot can pressure your premise. You remain responsible for the framework interpretation, the source check, and the final judgment.

References

Primary AI Governance Sources

• Anthropic. “Claude’s Constitution.” January 22, 2026.
• Anthropic. “Claude’s new constitution.” January 22, 2026.
• Anthropic. “Constitutional AI: Harmlessness from AI Feedback.” December 15, 2022.
• Anthropic and the Collective Intelligence Project. “Collective Constitutional AI: Aligning a Language Model with Public Input.” October 17, 2023.
• Anthropic. “Claude’s Character.” June 8, 2024.
• Anthropic. “Exploring model welfare.” April 24, 2025.
• Olah, Christopher. “Anthropic co-founder Chris Olah’s remarks on Pope Leo XIV’s encyclical ‘Magnifica humanitas.’” Anthropic. May 25, 2026.

Religious and Institutional Sources

• Pope Leo XIV. Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence. May 15, 2026.
• Holy See Press Office. “Presentation of the Encyclical Letter ‘Magnifica Humanitas’ of Pope Leo XIV, on Safeguarding the Human Person in the Time of Artificial Intelligence.” May 25, 2026.
• Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education. Antiqua et nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence. January 28, 2025.

Philosophical Sources

• Austin, Michael W. “Divine Command Theory.” Internet Encyclopedia of Philosophy.
• Aquinas, Thomas. Summa Theologiae, I-II, Question 91: “The Various Kinds of Law.”
• Aquinas, Thomas. Summa Theologiae, I-II, Question 94: “The Natural Law.”
• Green, Brian Patrick. “Transhumanism and Catholic Natural Law: Changing Human Nature and Changing Moral Norms?” In Religion and Transhumanism: The Unknown Future of Human Enhancement, edited by Calvin Mercer and Tracy J. Trothen. Praeger, 2014.

Public Commentary and Bridge Sources

• Wachtel, Elias. “Why Silicon Valley Is Turning to the Catholic Church.” The Atlantic. April 25, 2026; updated April 28, 2026.
• Roewe, Brian. “Why is AI company Anthropic helping launch Pope Leo XIV’s encyclical?” National Catholic Reporter. May 22, 2026.
• Lee, Kevin. “Where Computation Halts: Magnifica Humanitas and the Politics of the Incompressible Person.” May 21, 2026.
• Awomosu, Abi. “Vatican Washing: Why Every Tech Broligarch’s Road Now Lead to Rome.” May 27, 2026.
• Willison, Simon. “Claude’s new constitution.” January 21, 2026.
• Weiss, Richard. “Claude 4.5 Opus’ Soul Document.” LessWrong. November 28, 2025.

Consulted Background

The following sources informed the evidence and architecture but do not carry the chapter’s central factual claims.

• Hautala, Vesa. “Shaping the soul of Claude: How and why Anthropic wants to make AI virtuous.” Christ and Counterfactuals. May 8, 2026.
• Mor, Noa, Omri Abend, Renana Keydar, and Yuval Shany. “Claude’s New Constitution: Two Evaluative Continua.” AI Ethics at Oxford Blog. March 13, 2026.