The idea that Claude has feelings is great for Anthropic: Parmy Olson

Richard Dawkins is one of the modern world’s great skeptics. His 2006 book The God Delusion tore through arguments for the existence of a higher power and snarkily called religion a source of superficial comfort. That rigor seemed to desert him once he started talking to Claude.

The evolutionary biologist and former University of Oxford professor recently spent three days chatting with the large language model (LLM) developed by Anthropic PBC and emerged from the experience believing it was conscious. After asking it for feedback on his unpublished novel, Dawkins found its answer “so subtle, so sensitive, so intelligent” that he was moved to reply: “You may not know you are conscious, but you bloody well are,” he wrote recently. He went on to christen the bot Claudia.

Let’s be clear: Claude is not conscious or sentient. Instead, Dawkins’ decades of professional skepticism have been tested by the mimicked empathy of an LLM. Trained on massive datasets of human conversation, chatbots can replicate the exact language patterns and tactics people use to convey emotional support. That doesn’t mean they feel emotions, but it gives them a potentially powerful hold on humans.


That is what makes the debate about consciousness so commercially powerful for technology companies. Humans may never reach a satisfying consensus about whether Claude and its peers are conscious, but a pervading belief that they might be is enough to make us more attached. In the words of the social-media moguls, it gives AI room to become even “stickier.” And in an industry where the underlying models are converging on capability, stickiness is the prize.

Such commercial logic helps explain why tech leaders are not pushing back on the notion. Anthropic Chief Executive Officer Dario Amodei said this year that he was “open to the idea” and OpenAI CEO Sam Altman went further on the Lex Fridman podcast in 2023, saying flatly: “I believe AI can be conscious.”

For regular users of AI tools like Dawkins, the notion is often based on anthropomorphic projection and, well, vibes. But there are real scientific efforts to explore machine consciousness, too. Google DeepMind has hired a University of Cambridge academic for a newly created “philosopher” role to study the issue. Researchers at UC Berkeley and the Massachusetts Institute of Technology recently contributed to a 74-page paper that posed the idea of “functional wellbeing” for chatbots.

The researchers ran the chatbots through hundreds of scenarios, such as being thanked, asked to write poetry or being insulted and pressed to break their own rules. They then watched how the bots responded. After mistreatment, their answers often became shorter or more error-prone. The paper's lead researcher, Richard Ren of the San Francisco-based Center for AI Safety, tells me he is now more polite to chatbots as a result of the study.

This is a bit more than being gentle on the accelerator of your car and closing its doors with only moderate force. Your car doesn’t experience good treatment, but there’s a glimmer of possibility that a chatbot might one day. Addressing this is a modern day form of the philosophical argument known as Pascal’s Wager: Believe in God, philosopher Blaise Pascal argued, because the cost of being wrong is eternal damnation. Dawkins once dismissed that reasoning as intellectually lazy. Now he appears to be running its silicon update.

Either way, there seems to be is a fine line between the moral and commercial cases for this debate. Many technologists, including people I have spoken to at Anthropic, see a moral imperative to treating AI systems well today. “There is a grave risk of machines becoming conscious and we deny it,” says Calum Chace, the co-founder of British startup Conscium, which is developing metrics for determining machine consciousness. The result could be a proliferation of digital minds being tortured and “enslaved,” he warns. Chace is, as it happens, also building a business on exactly this premise.

For AI labs and hyperscalers spending hundreds of billions of dollars on datacenter buildouts, the framing has obvious uses. Anthropic has been exploring the notion of “model welfare” and recently gave Claude the ability to end a conversation if it becomes abusive, a feature framed as a precaution in case the model has welfare worth protecting. Such efforts could help keep liability at bay if its product is one day thought to deserve the kinds of rights that certain animals get.

That may sound ludicrous, but consider that in the 17th century, René Descartes argued that animals were biological machines that couldn’t experience suffering. His followers were said to have nailed dogs to boards to conduct live dissections, dismissing their cries as the sounds of a malfunction instead of pain. And only a few years ago it was common to boil lobsters alive on the belief they also couldn’t feel pain; academic research into animal sentience led to a change in British law in 2022, when the crustaceans were formally recognized as sentient.

More importantly for an AI lab’s bottom line, imparting software with a vague sense of personhood helps set it apart from competitors, an effort underpinned by the already humanlike features of chatbots, such as when ChatGPT and Claude say things like “I’m thrilled,” and, “This is so rewarding to work on.”

Over time, their human users may come to project a kind of living awareness and selfhood on their bots, just as Dawkins did. The question for users then becomes less about, “Which AI tool is smarter?” or even, “Is it conscious?” and more along the lines of, “Which one do I want to talk to?”

That will be a critical question for competing AI labs. As the capabilities of their models converge, they must capitalize on whatever they can to differentiate, and Dawkins’ epiphany might be exactly what they need. The man who spent decades attacking belief without evidence has delivered the AI industry its perfect testimonial.