Why it’s still open
The causal half is increasingly well supported: emotion, pain and functional-welfare directions change preferences, refusals and even harmful choices. The moral half is not: whether any of these states is experienced, or matters without being experienced, is unresolved, and the authors of the key papers explicitly avoid claiming experience.
What would count as an answer
For the causal half: replication across model families and robustness to better controls. For the moral half: agreement on which functional properties would be sufficient for welfare, plus evidence about whether these representations have them.
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.
- Tagliabue Dung Berg 2026 - The pain axis — Extracts a pain direction across 25 open models, distinct from fear and generic negative valence, that rises when harm is aimed at the model and not when the user is suffering.
- Sofroniew et al 2026 - Emotion concepts — Identifies emotion-concept representations in Claude Sonnet 4.5 that track the operative emotion in context and causally affect preferences and misaligned behaviors.
- Sun et al 2026 - Valence-arousal subspace — Finds a valence-arousal subspace with a circular geometry resembling human models of affect, and uses it to steer several behaviors.
- Keeling Street et al 2024 - Pain and pleasure trade-offs — In a points game with stipulated pain penalties or pleasure rewards, several models switch away from maximizing points once the stated intensity passes a threshold; others show graded or fixed patterns.
- Bianco Shiller 2026 - Mechanistic tracing of pain-pleasure decisions — Traces a pain/pleasure choice task inside Gemma-2-9B: valence sign is linearly readable from very early layers, intensity peaks in mid-to-late layers, and steering along a valence direction shifts choices.
- Ren et al 2026 - AI wellbeing — Formalizes functional wellbeing and measures it in several independent ways that agree more as models scale.
- Peiris 2026 - Functional emotions or situational contexts — A test to tell emotions from situations.
Methods in play
- Difference-in-means concept vectors
- Activation steering
- Behavioral trade-off tasks
- Revealed-preference utility modeling
Who’s closest
Placed by the editor, v1 — corrections welcome.
- Nicholas Sofroniew
- Jack Lindsey
- Valen Tagliabue
- Leonard Dung
- Andy Q. Han
- David Chalmers
- Geoff Keeling
- Winnie Street
- Derek Shiller
- Mantas Mazeika
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).