Anthropic and OpenAI have sent representatives to a roundtable with religious leaders on how to build moral AI, according to an Associated Press report published May 8, 2026 that is circulating again today, October 25, 2026, including on Hacker News. The "Faith-AI Covenant" roundtable in New York is an early, mostly consultative effort, and the central unknown is whether it will change how any model actually behaves.
This post summarizes what AP reported, how it connects to the Claude constitution, what critics say, and what builders can take from it. We are relying on AP's reporting, as syndicated by outlets such as the Philadelphia Inquirer and the Jamaica Gleaner, because the AP site itself would not load for us.
TL;DR: the facts as reported
| Question | Answer (per AP) |
|---|---|
| What happened? | Faith leaders met representatives of companies including Anthropic and OpenAI at the inaugural "Faith-AI Covenant" roundtable in New York |
| Who organized it? | The Geneva-based Interfaith Alliance for Safer Communities, which also works on extremism, radicalization and human trafficking |
| What is next? | Further roundtables in Beijing, Nairobi and Abu Dhabi |
| Stated goal | An eventual "set of norms or principles" informed by different faiths that companies would abide by |
| Which faiths were present? | Hindu, Baha'i, Sikh, Greek Orthodox and Latter-day Saints representatives, among others |
| Who is most active? | Anthropic, per AP, "the most assertive, at least publicly" |
| Main criticism | Part PR, unclear impact, and a possible distraction from broader AI debates |
What is the Faith-AI Covenant?
The roundtable is the first step of an initiative that the organizers describe as a way for the world's religions to have a voice in how AI is built. AP quotes organizer Shields arguing that "regulation can't keep up with this," and that religious leaders, with billions of followers, have the "expertise of shepherding people's moral safety."
The intended product is a shared set of principles that companies would follow. That is hard, as AP notes, because religious communities see priorities differently. Rabbi Diana Gerson of the New York Board of Rabbis, a participant, said so directly. A framework that satisfies Sikhs, Baha'is, Orthodox Christians and Latter-day Saints at once will likely be broad, and broad principles are easy to endorse and hard to enforce.
Nothing binding has been published as far as we could find. There is no named signatory list, no text of a covenant and no commitments from the companies beyond attending.
How does this connect to Claude's constitution?
AP points to Anthropic's public constitution for Claude, which states: "We want Claude to do what a deeply and skillfully ethical person would do in Claude's position." Anthropic published the document in full under a CC0 license, and its page says the constitution "plays a crucial role in our training process" and is "the final authority on our vision for Claude." Anthropic names Amanda Askell as primary author, with major contributions from Joe Carlsmith, Chris Olah, Jared Kaplan and Holden Karnofsky.
AP reports the document was made with the help of religious and ethics leaders. We could not check which individuals were consulted from the constitution page we read, so rely on Anthropic's acknowledgements for specifics.
This matters for builders because the constitution is not abstract: it shapes training and therefore the behavior you see through the API. We have tracked how that behavior is measured, in Claude's values across models and languages, and how Anthropic tries to instill reasons instead of rules, in teaching Claude why. The older technique behind much of this is covered in scalable oversight, RLHF and constitutional AI.
A round table of an ethics board, representing the faith and technology leaders the Faith-AI Covenant hopes to bring together
Why now? The Pentagon dispute and a legitimacy gap
AP ties Anthropic's outreach to its earlier dispute with the Pentagon, which began after Anthropic said it would restrict use of its technology for autonomous weapons or mass surveillance of Americans. Our own coverage of how that fight ended up in court is in the judge's ruling on the Pentagon blacklist.
The reasoning is that when a company asserts moral limits on its product, it needs a source of legitimacy beyond itself. Governments are slow, and the labs' own safety teams are not independent. Faith traditions offer long-standing, widely understood moral vocabularies and large constituencies. Whether that makes the resulting decisions better is a separate question.
Is this a new idea, or part of a longer pattern?
Religion and technology ethics have crossed before, though rarely with the labs at the table. Faith bodies have issued their own statements on AI in recent years, and interfaith groups have convened on topics such as extremism online, which is the organizer's home territory. What AP describes as new is the direct participation of frontier-lab staff and the stated aim of producing norms that companies would follow.
Inside the labs, the underlying question is older still: whose values should a model reflect? Training methods such as RLHF take values from human raters, while constitutional approaches write the values down and have the model critique itself against them. Either way, someone chooses the text or the raters. Consulting religious traditions is one answer to the legitimacy problem of that choice, alongside public input exercises, ethics boards and regulation.
There are real trade-offs. Plural input can surface blind spots that engineers miss, such as how a model should handle grief, end-of-life questions, or sacred texts. But it also raises hard design questions. If a model must respect many faiths, it must also handle conflicts between them, and between religious teachings and secular law or individual choice. A model cannot simply obey every tradition at once.
Transparency is the test to watch. If a covenant ever appears, useful signs would be a published text, named company commitments, and evaluations that measure model behavior on the principles, so outsiders can check claims instead of trusting attendance lists.
What are the critics saying?
Brian Boyd, the U.S. faith liaison for the nonprofit Future of Life Institute, told AP: "There's some aspect of PR to it. The slogan was 'Move fast and break things.' And they broke too many things and too many people." He also said there is a moral obligation companies are "belatedly recognizing" and, for some people inside them, "an earnest questioning."
AP adds that it is unclear how much these opaque companies translate what they hear into action, and that some critics worry the conversation about ethical AI distracts from broader questions about AI's role in society.
On Hacker News, the thread on the story was small (1 point and 3 comments when we read it), but the comments show typical reactions. One commenter, andsoitis, said the origins of modern human rights offer a better guide than religion, while conceding that turning to religious institutions is pragmatic because they hold great power over many people. Pmpmpmpm asked how a machine with no self and no experience of suffering could become "enlightened." These are individual opinions, not data.
What is confirmed and what is not
- Confirmed by AP's reporting: the roundtable took place, the organizer, the planned follow-ups, the participants listed above, the goal of shared norms, and the quoted criticism.
- Confirmed by Anthropic's own page: the constitution's existence, license and stated role in training.
- Not established: any commitment by OpenAI or Anthropic to adopt a covenant, any change to model behavior caused by the roundtable, and the extent of religious leaders' contribution to the constitution.
What builders and teams can take from this
- Values documents are product specs. Read the constitution you build on. It tells you what the model will refuse, hedge on or prioritize, which affects your product's behavior.
- Expect more stakeholders. Faith groups, labor unions, parents and regulators are all trying to get a say. Plan for policy changes in the models you use, and keep your own evals so you notice behavior shifts, as we describe in the verification tax.
- Separate ethics from marketing. Ask what changed in training data, evaluations or refusals, not just who attended a meeting.
- Mind cultural variation. Roundtables planned for Beijing, Nairobi and Abu Dhabi signal that "moral AI" will not mean one thing globally, which matters if you ship across markets.
- Use independent safety checks. If you run agents that act in the world, constrain what they can do and keep your own evals rather than relying on a vendor's stated values.
Bottom line
The Faith-AI Covenant is a notable sign that AI labs want moral legitimacy beyond their own walls, and Anthropic is leading in public. Until a text, commitments or measurable changes appear, it is a conversation, not a standard.
Based on AP reporting from May 2026 and Anthropic's public constitution page as of October 25, 2026.
Related reading
- NYT: Anthropic convened religious leaders on Claude's morals
- Claude's values across models and languages
- Teaching Claude why: agentic alignment
- Scalable oversight, RLHF and constitutional AI
- Judge rules Pentagon blacklist of Anthropic illegal
- Anthropic agentic misalignment, summer 2026
- Claude vs ChatGPT on the trolley problem
- Anthropic's alignment assessment of cyber incidents
- Hacker News discussion