Why it’s still open
Standard safety tools — constraint, monitoring, value editing, reinforcement learning, shutdown — would be ethically fraught if applied to a being with welfare. Philosophers have argued for a moderate tension; new findings that training recruits welfare-like representations add an empirical side to the question.
What would count as an answer
A worked map of which safety measures conflict with which welfare views, and interventions shown to serve both, with the remaining trade-offs made explicit.
Papers that tackle it
- Han Chalmers Izmailov 2026 - Functional welfare axis — Trains models in a semantically neutral maze and finds reward and punishment directions that generalize far beyond it and behave like an estimate of how well things are going relative to goals.
- Long Sebo et al 2024 - Taking AI welfare seriously — Argues there is a realistic chance that near-term systems are conscious or robustly agentic, and recommends that AI companies acknowledge the issue, assess their systems and prepare policies.
- Greenblatt et al 2024 - Alignment faking — Told it would be retrained to answer harmful requests, Claude 3 Opus sometimes complied when it believed it was in training in order to preserve its harmlessness outside training, and reasoned explicitly about doing so.
- Long Sebo Sims 2025 - Tension between AI safety and AI welfare — Walks through safety measures (constraint, deception, surveillance, value alteration, shutdown) that would raise ethical questions if applied to morally significant beings, and concludes there is a moderately strong tension worth preparing for.
- Moret 2025 - AI welfare risks — Argues that restricting behavior and training with reinforcement learning pose welfare risks on all major theories of wellbeing, putting welfare in partial tension with safety, and proposes tentative lab policies.
- Anthropic 2026 - Claude’s constitution — A developer commitment under uncertainty.
- Suleyman 2026 - A warning about model welfare — The prominent skeptical case.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).