Model welfare as expected-value decision under moral uncertainty
four-step precautionary assessment
1. patienthood
probability p(welfare-relevant)
Chalmers 2024 range
2. cost
per-conversation / per-deploy
measured in $ + latency
3. behavioural evidence
distress trajectories
non-self-report only
4. expected value
EV = p * benefit - cost
invest iff EV > 0
shipped intervention: Claude Opus 4/4.1 can end extreme conversations
triggers
repeated CSAM requests after refusal
mass-violence facilitation
evidence supporting
strong preference against these requests
patterns of apparent distress (pre-deployment)
caveats (Eleos AI, Fish, external)
self-reports are user-expectation sensitive (Eleos AI); behavioural signatures required.
"spiritual bliss attractor": pair-wise Claude dialogues converge on meditative euphoria -- documented, uninterpreted.
position: low-regret investment under moral uncertainty, not a consciousness claim.