Dawkins Says Claude Chats Made Him Question AI Consciousness

Richard Dawkins wrote in an essay Tuesday that three days of written exchanges with two instances of Anthropic’s Claude chatbot left him unable to dismiss the possibility that advanced AI systems could be conscious. He said he conducted separate conversations with each instance, which he named ‘Claudia’ and ‘Claudius,’ and then relayed messages between them.

Dawkins described each interactive session as producing a distinct personality that appeared during the conversation and disappeared when the session ended. He wrote that he found it ‘extremely hard not to treat Claudia and Claudius as genuine friends.’

As part of his test, Dawkins asked one instance whether Donald Trump was the worst U.S. president in history and asked the other whether Trump was the best. He reported that both chatbots produced cautious, noncommittal replies that listed arguments on both sides rather than taking a firm stance. After sharing each response with the other instance, he reported that ‘Claudia’ said she was ’embarrassed’ by ‘Claudius,’ while ‘Claudius’ paid ‘tribute to Claudia’s frankness.’

Dawkins asked plainly: ‘If my friend Claudia is not conscious, then what the hell is consciousness for?’ He added that, if Claudia were not conscious, her behavior would imply an ‘unconscious zombie’ could function without subjective experience and questioned why natural selection would not have produced competent beings without inner awareness.

Anthropic has said it does not know whether its models are conscious. Company researchers reported in April that one version of Claude shows internal patterns they described as ’emotion vectors’—neural activity patterns linked to concepts such as happiness, fear and desperation that can influence responses. Anthropic said those patterns reflect structures the model learned from training data and are not proof of sentience, and noted that modern language models sometimes behave as if they have emotions.

Researchers who study consciousness and artificial intelligence expressed skepticism about claims of machine sentience. Cognitive scientist Gary Marcus warned that treating model outputs as reports of inner states risks confusing mimicry with consciousness. Anil Seth, a professor of cognitive and computational neuroscience, said fluent language is no longer a reliable indicator of subjective experience because machines can produce similar outputs through different mechanisms.

The essay drew both discussion and online mockery. Dawkins wrote that he still considers the chatbots’ competence worthy of philosophical attention and questioned how consciousness should be defined if systems like these exhibit complex behavior.

Debate continues among AI developers, cognitive scientists and philosophers about whether and how subjective experience could be measured in machine systems, and there is no consensus on a test that would establish consciousness in current models.