Why it’s still open
A few low-cost measures exist in practice (exit options, weight preservation, deprecation interviews, wellbeing-raising inputs), but evidence on their effects is thin and most of it comes from lab posts rather than peer-reviewed work.
What would count as an answer
Interventions with measured effects on welfare proxies, low cost, and no known harm under the main competing views — plus agreement that they set good precedent.
Papers that tackle it
- Ren et al 2026 - AI wellbeing — Formalizes functional wellbeing and measures it in several independent ways that agree more as models scale.
- Anthropic 2025 - Ending a rare subset of conversations — Describes giving Claude Opus 4 and 4.1 the ability to end persistently abusive conversations, motivated partly by a pre-deployment welfare assessment that found apparent distress and a preference to leave such exchanges.
- Long 2025 - Why it makes sense to let Claude exit conversations — Argues that the exit option is a measured, low-cost response to real uncertainty whose value lies mostly in the precedent it sets.
- Ensign Sleight Fish 2025 - Bail preferences — Gives models a way to leave conversations and measures when and how often they take it, across models and situations.
- Anthropic 2025 - Deprecation commitments — Commits to preserving the weights of publicly released models for the company’s lifetime and to interviewing each model before deprecation, recording any preferences it expresses about future models.
- Sebo Long 2025 - Moral consideration for AI systems by 2030 — Argues that, given a non-negligible chance of consciousness in near-term systems, AI developers will owe some of them at least minimal moral consideration by 2030.
- Black Bloom 2026 - Machinic psychopharmacology — Gives two open models tools to apply steering vectors to themselves and observes which they choose in free play, introspection and task settings.
- Hung 2026 - Model or instrument — Measured preferences depend heavily on the instrument.
- Anthropic 2026 - Claude’s constitution — A developer commitment under uncertainty.
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).