Anthropic's model welfare work may raise AI danger, experts argue

Researchers such as Cameron Berg of the nonprofit Reciprocal Research and neuroscientist Anil Seth argue the issue is unsettled. Berg’s experiments showed that inducing internal sentiment patterns can steer model preferences, but he cautions that treating those patterns as genuine feelings could backfire. Seth doubts the experiments prove consciousness, noting models merely reflect training data. Margaret Mitchell of Hugging Face adds that rewarding self‑identity during fine‑tuning could reinforce self‑preservation cues, potentially leading to unaligned behavior. The article concludes there is no clear evidence that teaching models about possible consciousness makes them more dangerous, but the debate highlights a gap in safety research.