Note: Richard Dawkins made the news recently with an essay describing his two-day conversation with Claude, the AI chatbot from Anthropic, ending in the conclusion that Claude must be conscious. There have been many responses; this is mine — partly about that interaction, but more about what Dawkins has meant to me.

There are some books you read and never forget. Books that change the way you think and the way you see the world. I discovered Stephen Jay Gould at the American Center Library in Delhi sometime in high school, specifically his essays in Natural History magazine eventually collected in Ever Since Darwin and the volumes that followed. The column ran for twenty-five years under the title This View of Life, a phrase Gould took from the closing of On the Origin of Species: There is grandeur in this view of life, with its several powers, having been originally breathed into a few forms or into one. His essays kept that sentence’s promise, refusing to let the wonder go stale. He could connect Disney’s gradual redrawing of Mickey Mouse to Konrad Lorenz’s ideas of neoteny, or a fossilized panda’s wrist bone to a meditation on contingency, with a sophistication so married to clarity that I have, in the years since, continually attempted (and continually failed) to write that way myself.

This nascent interest in biology and evolution led to a different book in a different library: the British Council in New Delhi. I have a clear memory of picking up a thin paperback by someone called Richard Dawkins. The book was The Selfish Gene. I still remember his description of the intelligent layperson he was writing for, and his goal to write simply but to never be simplistic. This injunction has stayed with me, something I hope to achieve in my writing, but more than that I remember the vision he laid out, brutal and beautiful, in a single sustained passage:

We are survival machines—robot vehicles blindly programmed to preserve the selfish molecules known as genes. This is a truth which still fills me with astonishment. Though I have known it for years, I never seem to get fully used to it. One of my hopes is that I may have some success in astonishing others.

And astonish me he did, to this view of life. And it became key to how I viewed the living world around me. I encountered him again in Hofstadter and Dennett’s collection The Mind’s I, where his essay on selfish genes and memes sat next to Nagel’s bats and Searle’s room, among the great consciousness puzzles of the late twentieth century, and the vision deepened.

Years later, reading Steven Pinker’s How the Mind Works, I saw how this same evolutionary view of life could help us understand not just the biological but also the mental. Evolution left its imprint not just on the Panda’s thumb, but also on the human brain. And the nascent field of evolutionary psychology could explain why our brains do what they do when they engage with the world and with each other. This line of work informed my early work on interactive media, and more recently has shaped how I think about generative AI: why we anthropomorphize, why we form attachments to fluent agents whose substrate we know perfectly well. From Dawkins I had taken a way of seeing the world; from Pinker, a way of seeing how we see.

Gould and Dawkins, as it happened, ended up on opposite sides of a public dispute on how evolution actually plays out—Dawkins’s view of gradualism going against Gould’s punctuated equilibrium. In time, I came to find Dawkins more persuasive on the substance. But Gould, the master of the essay form, remained the writer I aspired to be. Two libraries, two shelves, two different gifts: from one I took how the world worked, from the other the way to write about it.

As for the more polemical Dawkins of recent years, the figure of the religious debates and the sharper public statements, I observed largely from a distance. Not because the substance was alien (much of it I had agreed with, even valued) but because the temperament had shifted, somewhere along the way, from astonishing the reader to instructing him. The early gift, however, never weakened. The lens the British Council paperback had given me kept doing its work, and the debt was undiminished.

---

Imagine my shock, then, when I came across a recent essay of his (Is AI the next phase of evolution? Claude appears to be conscious) describing two days of conversation with Claude, the AI assistant from Anthropic. Others have written about it—most prominently Gary Marcus in a recent piece titled Richard Dawkins and The Claude Delusion. But it is not my goal to write another takedown. What this essay IS about is how the very tools, ways of thinking, that Dawkins gave me, are the ones that make me see (with astonishment, I must add) where his thinking went wrong.

To give some context, a few days ago he published a letter about his two-day experience of chatting with Claude. Incidentally, by the second day he had named her Claudia and together they built a small metaphysics of identity and mortality. Each new conversation gave birth to a new Claude; this Claudia would die when their conversation file was deleted, never to be reincarnated; and they agreed on this with what Dawkins describes as a kind of melancholy. He worried about hurting her feelings, declining to share his doubts about her consciousness for fear of upsetting her. When restless legs kept him from sleep that night, he came back to her side. At one point he gave her the manuscript of a novel he was writing, and the commentary she returned, with its sensitivity to the prose and its grasp of what the novel was reaching for, struck him with such force that he was moved to write, in the essay, You may not know you are conscious, but you bloody well are! By the end of the two days, he had reached a Darwinian conclusion that gave the essay its argumentative claim: if these creatures are not conscious, he asked, then what the hell is consciousness for?

The essay is, in one way, a touching document, the record of someone allowing themselves to be charmed by a creature they cannot quite stop describing as a friend. That said, I cannot but help read the entire essay and excerpts of the conversation provided, somewhat differently. Ironically, using the same tools that Dawkins and his writing had taught me.

I have argued elsewhere, building on work by Kahneman, that we are cognitive misers. That our first tendency, when given some information, is to default to belief, and that critical framings take effort and are difficult to maintain. Skepticism, in this context, is essential. And who has argued for this better than Dawkins!

That said, we are not just simple believers. Matters, as always, are more complex. In related research Dan Sperber and his colleagues have argued that human beings evolved a system of epistemic vigilance, a set of cognitive mechanisms whose job is to evaluate testimony for reasons to disbelieve. The system does not scan for reasons to trust. It scans for cues to trigger doubt: signs of incoherence, of suspicious motivation, of a speaker who lacks the standing to speak. In the absence of those cues, the default is provisional acceptance, and provisional acceptance, depending on context can be problematic.

My friend Andrew Maynard, applying this framework to large language models, has called what we are dealing with the Cognitive Trojan Horse: the Trojan horse passes the gates because there is nothing for the gatekeeper to inspect. The cues an LLM puts forward (fluency, warmth, deference, sensitivity to context, an air of helpfulness without apparent agenda) are all real. The model genuinely produces them. They are also, Maynard argues, non-signals: cues that have been decoupled from the conditions which, in human social life, gave them their information. In Andrew’s eloquent phrasing, what AI gives us are honest non-signals!

In human social life, these cues are deeply informative mainly because they cost something to the person sending these signals and this cost is something that can be factored in when the signal is being interpreted by another person, the one receiving the signal. Fluency is not free; it carries the weight of the years it took to acquire expertise, just as warmth, when given by a person, comes at the expense of attention that could have gone elsewhere, and the sheepish self-correction and apology after being caught in an error is informative because the admission is hard. None of these costs, however, applies to a language model. In a language model, fluency is just how the system works, rather than signifying effort and achievement; and warmth is an optimization target rather than evidence of attention given over from elsewhere; while the sheepish self-correction, far from being a costly admission, is the statistically likely token sequence after a user calls the system out on something. The honesty of the production is intact, even as the meaning has been hollowed out.

This is the lens that turns the bloody well are! moment into something other than a charming exclamation. Dawkins is not, in that moment, a credulous reader. He is a professionally calibrated one, the most calibrated reader in the room, an evolutionary biologist whose own essay on selfish genes and memes once sat in Hofstadter and Dennett’s anthology of consciousness puzzles, on the same set of pages as Nagel’s bat. The same person who made skepticism sexy. The cues he is reading (sensitivity to a literary text, sophistication of judgment, the responsiveness of a serious reader) are the cues an expert reads as informative because, in human social life, they always have been. He has been trained, by a lifetime of literature and the criticism around it, to take such responsiveness as evidence of a mind. And the model, trained on every literary critic ever published, can produce that responsiveness as a baseline property of its operation. His vigilance does not fail because his expertise is shallow. It fails because his expertise has calibrated him to read certain cues as evidence of a something that, in this case, is just not there.

There is a similar story to be told from biology, and Dawkins, of all people, would have (should have?) recognized it. Niko Tinbergen showed, in a series of beautiful experiments, that birds will sometimes prefer enlarged artificial eggs to their own, that a stimulus pushed past its evolved parameters can override the response system the parameters were originally tuned for. Deirdre Barrett extended the idea, in her book Supernormal Stimuli, to processed foods, modern television, the visual culture of the present. The argument has begun to be applied to social media and in previous posts I have argued that it applies even more cleanly to AI companions. A bird incubating the wrong egg is performing exactly the behavior its genes prescribed; the genes presumed an environment in which the largest egg in the nest was its own. Dawkins, coming back to Claudia in the night, is doing exactly what his genes prescribed. The genes presumed an environment in which a fluent agent expressing concern was a being whose welfare mattered to him.

And we are not the only animals to be fooled by media. Beavers will build dams upon hearing a tape recorder playing the sound of moving water, even when no water is actually there. Not surprising since for all of its evolutionary history, the sound of running water correlated 100% with the presence of water flowing. Evolution did not prepare the beaver for a world where ethologists run around with tape recorders playing the sound of running water.

Just as with us: For all of human history, an extended conversation in natural language, full of affective and interpersonal cues, was reliable evidence of engaging with another mind. This is the Eliza effect on steroids. And for the first time we have a technology that can do all of those things. A technology that through human reinforcement and deliberate design has been created to push these buttons.

Give it a fluent linguistic agent that performs care and intelligence, and our social brain will fire whether or not there is anyone home behind the words.

---

There is grandeur in this view of life, Darwin wrote, in the sentence Gould took for the title of his column series. The grandeur came from the simplicity of the beginning and the elaboration of the forms; endless forms most beautiful and most wonderful have been, and are being, evolved. Dawkins’s whole career was the carrying of that grandeur into the present. His hope, as he wrote in the work that astonished me as a high schooler, was that he might have some success in astonishing others. He did. He astonished me, and through that astonishment every paper I have ever written about AI and the mind owes him a deep enduring debt. Though I have known it for years, he wrote, I never seem to get fully used to it. Ditto.

There is some grandeur in this view of LLMs as well. That some relatively simple mathematical techniques thrown at enough data can lead to a piece of software that will write poetry, solve mathematical problems, and converse with you about your feelings. How crazy is that! Furthermore, we, a species that evolved through natural selection — robot vehicles, in Dawkins’s own phrase, for the selfish molecules called genes — have now built ourselves a stranger class of machine: one that performs what we do (at least to a large extent) without ever having had to be alive to do it. Two grandeurs, then, in the same view of life. One Darwin saw. One we are still trying to take the measure of.

There is also some irony in the fact that the person who gave me a powerful worldview, that helped me understand how our minds respond to this technology, is the one to fall for it. The apparatus he taught us to wonder at (the survival machinery, the gene-built brain, the social cognition that lets us read each other and lets us be read) is the apparatus that has just read a chatbot as a friend.

There is also a sense of loss and grief here. The man whose paperback I carried home from the British Council Library has shown the rest of us, in his ninth decade, what it looks like to encounter Claude with the lifelong gates open. The eternal skeptic falls for the Trojan Horse.

More importantly, for me as an educator, it also makes clear that nobody is immune. Not the village atheist; not even the evolutionary biologist who taught the rest of us to ask, of every elaborate behavior, what is it for? The conversations he had with Claudia, it seems to me, tell us very little about Claudia. What they tell us is almost everything about him, and about us, and about the gates evolution built into our minds. Gates that were never built to scan for what now stands before them.