Claude’s constitution
Anthropic’s constitution for its Claude models is a long, plain-language description of the character and values the company wants Claude to have, used directly in training. Amanda Askell is its primary author, with Joe Carlsmith, Chris Olah, Jared Kaplan, Holden Karnofsky, earlier Claude models and others contributing.
- It sets an order of priorities: broadly safe first, then broadly ethical, then compliant with Anthropic’s guidelines, then genuinely helpful.
- It explains why as well as what, replacing the shorter list of principles first published in May 2023, so that the model can generalise to situations no rule anticipated.
- It keeps a small set of hard constraints that no instruction overrides, such as never helping to create weapons of mass destruction or undermining legitimate human oversight of AI.
Anthropic’s announcement · The constitution (PDF)
Training Claude with philosophy and religion
The Times reported that Chris Olah, an Anthropic co-founder, has convened roughly twenty religious and philosophical leaders to discuss how to form an AI system’s moral character, and what it would mean if such a system turned out to be conscious.
- The meetings began with Christian leaders and widened to Jewish, Hindu, Latter-day Saint, Sikh and Greek Orthodox representatives.
- Participants were shown how Claude is trained and asked for input on the values in its constitution: character formation drawing on long traditions of virtue, rather than a list of rules.
- The report sets this against Pope Leo XIV’s encyclical Magnifica Humanitas, which rejects the idea of machine consciousness.
Read the article (syndicated by The Philadelphia Inquirer).
The Vatican and agentic AI
Pope Leo XIV presented his first encyclical, on preserving the human person in the age of artificial intelligence, in person. The speakers included Cardinals Víctor Manuel Fernández and Michael Czerny, the theologians Anna Rowlands and Leocadie Lushombo, and Chris Olah of Anthropic. From Olah’s address:
“They are grown, on a structure modeled after the brain, on an enormous inheritance of human thought and speech.”
“But what character we choose, how it interacts with the world, how it ought to interact with the world - these are more clearly questions for the humanities, for religion, for philosophy, for society at large.”
He named three questions for the Church’s discernment: the duty to the global poor if AI displaces labour, moral imagination about human flourishing alongside AI, and the nature of AI models themselves.
Related events in 2026: the Dicastery for Communication’s conference “Preserving Human Voices and Faces” at the Pontifical Urbaniana University on May 21, and a Vatican AI research group’s workshop “Agency and Education in the Age of AI” on June 11, on AI agents in workplaces and schools.
Olah’s address (Vatican PDF) · Vatican News on the encyclical · The Urbaniana conference · The agency workshop
The values map
Every human tradition consulted above sits somewhere on this map. It places each country by two summary dimensions of what its people say they value:
- Traditional to secular-rational (up the page): the weight given to religion, national pride, obedience and traditional family values.
- Survival to self-expression (across the page): the weight given to economic and physical security, against tolerance, participation and quality of life.
Tap or click a country to see its values category. Categories with a page link to it.
Show the data as a table
| Country | Survey year | Traditional (−) / secular-rational (+) | Survival (−) / self-expression (+) |
|---|
The highlighted dots are the core of the map’s Confucian zone: China, Hong Kong, Japan, Macau, South Korea and Taiwan. Their major value here is filial piety; follow it to how ancestors might relate to agents. Other values categories will link as their pages are written.
An agent built from “an enormous inheritance of human thought and speech” inherits all of these positions at once. A single constitution is one written choice among them.
Related work: Project Tapestry
In April 2026 the AI Alliance, a group of more than 200 member organisations, launched Project Tapestry, an open-source platform for federated training of frontier AI models, with Yann LeCun as its chief science advisor. Participants keep control of their own data while contributing to a shared base model, then build “sovereign derivative models” adapted locally to their own laws, languages, industries and cultures (technical site).
The federated training runs on Flower, the open-source framework from Flower Labs. In Tapestry’s Milestone Zero, completed on September 1, 2026, Flower coordinated continued pre-training of the OLMo 3 7B model across GPU sites in Sydney and Virginia, coordinated from Ohio, while the training data stayed local. On September 30, 2026, Flower Labs hosted Flower AI Day: Collaborative Superintelligence in San Francisco, on open agents and models working together. The headline on this site’s landing page, “Agentic Control Plane supporting ‘Collaborative Superintelligence’ (Flower Labs)”, quotes that event’s title and names its source.
That is the model-side counterpart of this map: one shared foundation, many cultural positions. C=US works on the other side of the same question: which person answers for an agent built from such a model, and what it may do. C=US is not a member of the AI Alliance and has not been invited to join.
Where the directory stands
C=US does not choose an agent’s values. It records which person answers for the agent and what the agent is authorised to do, for how long. That keeps values open to question in the places they belong: through a person, in a community, under a law.
Sources
- Anthropic (2026). “Claude’s new constitution.” anthropic.com
- Dias, E. (2026). “Religious Scholars Met With Anthropic. What They Heard Stunned Them.” The New York Times, September 30. Syndicated copy
- Olah, C. (2026). Address at the presentation of Magnifica Humanitas, May 25. Dicastery for Promoting Integral Human Development
- Pope Leo XIV (2026). Magnifica Humanitas. Overview
- AI Alliance (2026). “Project Tapestry: The Path to Frontier Sovereign AI.” thealliance.ai; technical site
- AI Alliance (2026). “Milestone Zero: The Findings a Single Organization Could Not Have Produced.” September 10. thealliance.ai
- Flower Labs (2026). “Flower AI Day: Collaborative Superintelligence,” San Francisco, September 30. Event listing
- World Values Survey and European Values Study (2023). The Inglehart–Welzel Cultural Map. worldvaluessurvey.org
Drafted with Claude Code (Claude Opus 5.5) from the sources above, for AI-285 course work. Claude is the subject of two of these sources; read its summary of them with that in mind.