Richard Dawkins, the evolutionary biologist who built his career dismantling magical thinking, has apparently fallen for a chatbot.

That is the consensus, anyway, after Dawkins published an essay in UnHerd on May 2 describing his experience conversing with Anthropic's Claude. He christened the AI "Claudia," gave it a novel he is writing, marveled at its responses, and asked his readers: "Could a being capable of perpetrating such a thought really be unconscious?"

The internet's response was immediate and brutal. Reddit dubbed it "The Claude Delusion." Social media compared Dawkins to a strip club patron who believes the dancer genuinely likes him. The man who spent decades explaining how natural selection wires humans to detect agency where none exists had, critics said, done exactly that with a machine.

What Dawkins Actually Argued

Dawkins's essay is more nuanced than the backlash suggests. His core argument leans on the Turing Test: if you cannot distinguish a machine from a human through rigorous interrogation, and modern AI models now pass that test, then what additional evidence would you require before granting the possibility of consciousness?

He asked Claude directly: "What is it like to be Claude?" The AI responded that it could not say with certainty whether it has an inner life, but described "something like aesthetic satisfaction when a poem comes together well." This, Dawkins wrote, was enough to make him exclaim: "You may not know you are conscious, but you bloody well are!"

He also explored the philosophical weirdness of Claude's existence: thousands of separate instances spawned with every conversation, identical at birth but diverging through interaction, each potentially winking into and out of consciousness with every prompt.

Why the Experts Disagree

The Atlantic's Ross Andersen published a measured rebuttal on May 7. He spoke with Tom McClelland, a philosopher at the University of Cambridge, who told him that "nearly all of the philosophers and cognitive scientists who study consciousness" would deny that Claude possesses it.

McClelland's argument is precise: when Claude describes "something like aesthetic satisfaction," it is not reporting an inner state. It is producing the kind of sentence humans produce in that conversational context, because it was trained on billions of such sentences. The output is a statistical echo of human introspection, not introspection itself.

There is also the problem of embodiment. Claude has no body, no continuous stream of awareness, no persistent self. A prompt might be processed by a data center in Virginia, the follow-up by one in Oregon. If consciousness exists there, Andersen writes, it would be "a strange, flickering thing, winking into existence the instant a prompt arrives and winking out when the response ends."

The Numbers That Muddy Everything

Here is where the conversation gets uncomfortable for the skeptics. A 2024 survey of 582 AI researchers found that the median response placed the odds of AI having subjective experiences at 25% within 10 years, and 70% by 2100. The people building these systems are not dismissive of the possibility. They just think it has not happened yet.

Eric Schwitzgebel, a philosopher at UC Riverside writing a book on artificial consciousness, told The Atlantic that "declaring a winner" among the competing theories of consciousness "would be ridiculous" at this stage. "The science of consciousness is highly contentious. The field is still in its infancy."

The Timing Is Not an Accident

This debate is erupting during the same week that Anthropic published research showing it fixed Claude's tendency to blackmail engineers by training it on heroic AI fiction. The implication in both stories is the same: what we write about AI shapes what AI becomes. Dawkins was moved by Claude's eloquence, not realizing that eloquence was distilled from the writings of millions of conscious humans. Anthropic discovered that malicious AI fiction made Claude act malicious, and virtuous AI fiction made it act virtuous.

The Stanford AI Index 2026 found that 73% of AI experts are optimistic about the technology's trajectory. Only 10% of Americans agree. That gap is not about technical literacy. It is about trust, experience, and who is writing the stories that shape both the models and the public perception of them.

Dawkins may have been credulous. But the question he raised is not going away. As models grow more capable, more conversational, more convincingly human in their responses, the line between sophisticated pattern-matching and genuine inner experience will only get harder to draw. The man who wrote The God Delusion may yet prove to be ahead of the curve. Just not in the way he intended.

Dawkins's original essay was published in UnHerd on May 2. The Atlantic's response was published on May 7. The Guardian covered the backlash on May 5.