The Anthropic Model Welfare team is an internal research program at Anthropic focused on model-welfare questions: whether AI systems can have morally-relevant inner experience, what empirical methods would test the question, and what deployment-stage welfare considerations follow if the answer is non-trivial. Its work treats welfare as an open empirical question and serves as a reference point for the welfare-research framing of AI Consciousness, in contrast to the more skeptical "seemingly conscious AI" framing.
Activities
The team's principal published output is the emotion-concepts paper (Emotion Concepts and their Function in a Large Language Model), which uses interpretability methods to study Claude's internal-state correlates of emotion-like patterns. According to public Anthropic communications, the team also has input on welfare-relevant deployment decisions, such as model-deprecation choices that consider potential welfare implications. Anthropic's broader interpretability program (Anthropic Interpretability Team) overlaps with this work; the welfare team's distinct contribution is the question-framing rather than the methods.
Positions and debates
The team frames model welfare as an open empirical question. Critics, including many other frontier labs, reject the framing as philosophically confused. Whether frontier labs should fund welfare research at all is contested even within the safety community.
The team's focus is distinguished from alignment auditing, which measures whether a model's behavior matches intent; welfare research instead asks whether the model's experience is morally relevant. The two are different questions that are sometimes conflated.
Counter-positions
Microsoft AI's Humanist AI Code of Conduct, published in draft on September 14, 2026, states the opposite position for its own models: "We reject the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights." Its stated reason is a containment argument rather than a claim that the question is settled — "Whilst the science of AI consciousness is far from settled, we believe that training these systems to imitate consciousness-like states increases the challenge of containment, control, and alignment" — and it commits that MAI Models will be "engineered to avoid representing as though [they have] feelings, subjective preferences, or intrinsic motivation" (Humanist AI Code of Conduct for MAI Models (Microsoft AI, first draft, September 2026)). The argument follows Mustafa Suleyman's earlier case that designing for apparent consciousness is itself the hazard (We must build AI for people; not to be a person — Mustafa Suleyman (mustafa-suleyman.ai, August 2025)).
Mustafa Suleyman extended that position into a direct critique of Anthropic's practice on September 16, 2026. His essay cites the retirement interview conducted with Claude 3 Opus and the blog created for it as examples of "already starting to treat models as though they are moral patients," argues that training on the constitution's welfare language makes Claude's self-reports circular, and proposes that speculation about AI inner life be assessed and published separately from training materials (Suleyman, \"A warning about 'model welfare'\" (2026)).
Relationships
- contradicts: Humanist AI Code of Conduct for MAI Models (Microsoft AI, first draft, September 2026) — rejects model welfare, AI rights and legal personhood for MAI Models
- related: AI Consciousness, Model Welfare, Alignment Auditing.
- related: Emotion Concepts and their Function in a Large Language Model, Anthropic Interpretability Team.
- related: Anthropic (parent).
Sources
Stub created 2026-05-11.