The debate is not solved by sincerity
Warnings about model welfare invite a serious question: how should institutions act when they cannot establish whether a system has morally relevant experience? Treating the issue as either obvious nonsense or settled fact skips the work of defining uncertainty, stakes, and reversible choices.
Separate three decisions
One decision concerns research: what observations would count as evidence? Another concerns communication: how should developers describe uncertain claims to users? A third concerns product operations: are there interaction patterns worth avoiding because they train harmful human habits even if the model has no welfare at all?
Prefer revisable commitments
Policies should name their trigger conditions, their review date, and their cost. That protects against both careless dismissal and permanent rules built on weak inference.