Why it’s still open
Every welfare or psychology claim needs a bearer, and the candidates come apart: one set of weights runs as many instances, a conversation can span model swaps, and one conversation can host several characters. Philosophers now defend different answers (virtual instance, thread, persona-based views), and the interpretability evidence that could decide between them — whether persona regions have real boundaries — is only starting to arrive.
What would count as an answer
A criterion that picks out one unit and survives hard cases: model swaps mid-conversation, drift into another persona, parallel copies. Ideally backed by mechanism, such as showing that the states that matter (preferences, valence, memory-like continuity) are carried by that unit and not another.
Papers that tackle it
- Chalmers 2025 - What we talk to — Argues that LLM interlocutors are quasi-agents with quasi-beliefs and quasi-desires, best identified with virtual instances or conversation threads rather than models or hardware, and draws out what this means for welfare.
- Beckmann Butlin 2026 - Where is the mind — Uses persona-vector research to widen the candidates for an LLM mind to three: the virtual instance, plus two new persona-based views.
- Register 2025 - Individuating artificial moral patients — Sets out the moral risks of not knowing how to count and re-identify AI moral patients (survival, counting, trade-offs, bodies) and explains why human personal-identity theories do not settle them.
- Birch 2025 - AI consciousness a centrist manifesto — Holds two challenges together: mass misattribution of human-like consciousness driven by mimicry and role-play, and the real possibility of alien consciousness that current theory cannot confidently rule in or out.
- Artificial Self 2026 - Landscape of AI identity — Maps the ways an AI system’s identity can be configured and how those configurations shape behavior, deliberately setting aside whether the system is conscious.
- Long Sebo et al 2026 - Studying AI welfare empirically — A framework: question × entity × evidence.
- Arbel Salib Goldstein 2026 - How to count AIs — A legal answer to counting AI agents.
- Lerchner 2026 - The abstraction fallacy — A substrate-skeptical argument from Google DeepMind.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
- Whose valence is it
- Counting moral patients
- Which activations belong to the Assistant
- What is owed at deprecation
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).