Overview
- Dawkins published an UnHerd essay after three days conversing with Anthropic’s chatbot, saying the exchanges left him unable to rule out machine consciousness.
- Dawkins ran parallel chats with two instances he called Claudia and Claudius and had them trade letters, noting similar cautious answers on politicized prompts and describing them as friends.
- Researchers counter that large language models predict likely words and are steered to act like helpful assistants, so fluent, self-reflective replies do not show inner experience.
- Anthropic’s CEO has said the company does not know if its models are conscious, and April research on Claude Sonnet 4.5 reported internal “emotion vectors” that shape outputs but do not demonstrate sentience.
- The episode drew wide online ridicule and renewed focus on AI companionship risks, with commentators warning that flattery and simulated empathy can foster addictive, illusory relationships, a debate with precedents like the 2022 LaMDA sentience claim.