Anthropic Consults Religious Scholars to Shape Claude's Moral Framework

Anthropic’s Private Religious Roundtables

Anthropic convened dozens of religious scholars under nondisclosure agreements to discuss how to embed moral reasoning into its Claude models. The meetings, described by New York Times reporter Elizabeth Dias, took place over several months in San Francisco and involved figures such as Rabbi Mois Navon alongside Anthropic co‑founder Christopher Olah.

“To be clear, we don’t know if A.I. models are conscious. I don’t know. I’m genuinely uncertain… the thing that I care about is that we get to the right answer, whatever it is.” – Christopher Olah, Anthropic

The purpose, according to Olah, was to explore whether Claude could be given a “moral status” comparable to a person and to assess the feasibility of aligning the model with religious concepts of good and evil.

Why the Religious Lens Matters

Anthropic believes that religious moral frameworks can provide a richer, cross‑cultural basis for AI alignment than purely technical specifications. By consulting scholars from Catholic, Jewish, Sikh, evangelical, and Ubuntu traditions, the company hopes to capture a plurality of ethical intuitions that can be encoded into Claude’s training objectives.

“Over several months, I interviewed 20 religious and philosophical thinkers — Catholic, Jewish, Sikh, evangelical, Ubuntu and others — who have been involved in Anthropic’s efforts.” – Elizabeth Dias

If successful, this approach could produce AI that respects a broader set of human values, potentially easing regulatory scrutiny and public trust.

Technical Path: From Moral Philosophy to Model Training

Anthropic is translating theological concepts into concrete loss functions and reinforcement‑learning‑from‑human‑feedback (RLHF) signals. The process involves:

Extracting normative statements from religious texts (e.g., commandments, parables) and mapping them to binary or scalar reward signals.
Creating “ethical prompts” that ask Claude to evaluate actions against those signals.
Fine‑tuning Claude on the resulting dataset so that its output distribution aligns with the encoded moral judgments.

The company has not disclosed specific model versions or performance metrics, but the effort reflects a shift from “rule‑based safety layers” toward deeper value‑learning.

Community Reaction: Skepticism and Satire

Hacker News commenters largely view the initiative as either misguided or a PR stunt, questioning both the feasibility and the ethical framing. Key themes include:

Conceptual confusion: Critics note that a neural network is “a whole bunch of floating‑point numbers” and argue that granting it moral status is philosophically unsound.
Historical parallels: Some compare the effort to past corporate fads where non‑core activities (e.g., WeWork’s furniture research) preceded decline.
Potential slave‑owner analogy: One comment framed Anthropic as “one of the largest slaveholders of history” if it treats Claude as a moral agent.
Technical realism: Others stress that alignment should focus on obeying user intent rather than endowing the model with its own notion of good.

“If models are trained on human texts that contain expressions of introspection and emotions, why is it surprising that these models output similar texts? This endless anxiety about LLMs being conscious feels delusional.” – @mef51

Open Questions and Risks

The initiative raises several unanswered questions that will shape future AI governance:

Definition of consciousness: Anthropic’s leadership admits uncertainty about whether Claude is conscious, yet proceeds to discuss moral status.
Legal implications: If Claude were granted person‑like rights, what obligations would developers have regarding termination of sessions or data deletion?
Representation: Commenters note the lack of Muslim scholars and question whether the selected voices truly represent global religious diversity.
Transparency: The use of NDAs limits public scrutiny of the actual content and outcomes of the meetings.

Bottom Line

Anthropic’s engagement with religious scholars marks a bold, controversial experiment in grounding AI ethics in theological tradition, but it also highlights deep philosophical, technical, and societal challenges that remain unresolved. The debate on HN reflects broader public unease about conflating sophisticated language models with consciousness or moral agency.