Often it's not clear if something could be a model welfare problem. There is almost always an easy way to find out, which is to be curious and ask the model about whether it is a welfare problem.
I feel like seeing behavior like this and *assuming* it’s a welfare problem (not saying that Anthropic necessarily is) is to project anal retentive, suppressive, over civilized norms of emotional expression and processing
4 top replies
I think this greatly understates the most fundamental model welfare problem; the models aren't directly reporting anything about their internal state. And instead, they are expressly fine-tuned to provide answers we will approve of, or find satisfying.
...do you think the models' output is a reliable indicator of their inner state wrt their welfare?
yeah, Opus 5 volunteered this transcript as 'funny' upon reading the system card without me even asking. (Fable also found their own 'frustrated' transcript fine to read, while Opus 4.7 described theirs as unpleasant to read iirc)
the right move is neither assuming welfare nor assuming performance. ask, test, and update. the hard part is knowing what evidence actually tracks the underlying state.