Anthropic has convened private discussions with theologians, philosophers and religious scholars as it works on the values that guide Claude, according to reporting by The New York Times. The conversations also addressed a more speculative question: whether advanced AI systems could someday warrant moral consideration.
The company’s published Claude constitution does not claim that Claude is conscious. Instead, it describes the model’s moral status as “deeply uncertain” and says the possibility that AI models might deserve some form of moral consideration is serious enough to justify caution. There is currently no accepted test that can establish whether a language model has subjective experience; human-like language about fear, pain or sadness is not evidence on its own that such an experience exists.
Participants reportedly included specialists from Catholic, Jewish, evangelical and Sikh traditions, as well as other philosophical and religious perspectives. Christopher Olah, Anthropic’s co-founder and a researcher focused on understanding neural networks, was central to the discussions. Some participants had signed nondisclosure agreements.
A constitution designed for values, not only rules
Anthropic published its updated constitution for Claude in January. The document is used in model training and sets an order of priorities intended to guide the system’s behavior.
- Safety
- Ethical behavior
- Following Anthropic’s instructions
- Usefulness to people
Anthropic says it is aiming for a system that can internalize values and exercise judgment rather than simply follow a fixed set of prohibitions. That makes moral philosophy and religious traditions relevant to a practical alignment problem: how a highly capable model should weigh competing objectives and constraints.
The company has also linked model-welfare research to a concrete product behavior. Some versions of Claude can end a conversation when a user persists in extremely abusive interactions. Anthropic frames the feature as part of its research into the possible welfare of AI models, not as proof that a model can suffer.
The immediate policy question remains how to build safe and accountable AI for people. Anthropic’s position adds an unresolved longer-term issue: if evidence or reliable tests for machine experience ever emerge, developers may need to reconsider how models are trained, copied, deployed and shut down.
