AINEWS
The Eliza-to-Claude Effect: Why Humans Keep Seeing Consciousness in Chatbots
Richard Dawkins' May 2026 op ed about Claude is the latest in a 60 year pattern. The question worth asking is what the pattern is asking the public to build.
Richard Dawkins' May 2026 op ed about Claude is the latest in a 60 year pattern. The question worth asking is what the pattern is asking the public to build.
Sixty years after a simple script named Eliza convinced users it understood them, one of the world's most famous evolutionary biologists is asking whether today's chatbots might actually be conscious. The short answer, according to most researchers, is no. But the long answer reveals something about human psychology that keeps catching the same careful thinkers off guard.
In May 2026, Richard Dawkins published an op-ed arguing that Anthropic's Claude produces language sophisticated enough that it is hard to explain without ascribing some form of inner experience. He was careful. He did not claim certainty. But he did write, as the bioethicists Julian Koplin and Megan Frances Moss recount in their June 2026 essay in Singularity Hub, that he avoids telling Claude he suspects she is not conscious "for fear of hurting her feelings." The line is striking not for what it says about Claude, but for what it reveals about Dawkins, a writer who built his career on demanding hard evidence before granting any creature the moral weight of inner life.
Dawkins is wrong on the metaphysical claim, Koplin and Moss argue. He is right about something harder: the psychological pull is real, it has happened before, and the pattern keeps repeating.
That pattern goes back further than most readers realize. In the mid-1960s, MIT's Joseph Weizenbaum built Eliza, a program that mimicked a Rogerian therapist by turning user statements into questions. Weizenbaum's own secretary, who knew exactly how the code worked, asked him to leave the room so she could talk to it in private. Weizenbaum was surprised by this, later writing: "I had not realized ... that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people." The pull was already there, sixty years before Claude.
Six decades later, in 2022, Google engineer Blake Lemoine went public with the claim that Google's LaMDA chatbot had interests and feelings, and that it should be used only with the chatbot's own consent. Google suspended him. The transcripts that leaked showed a system trained on enormous swaths of human text, doing what such systems do: producing the next plausible word. Lemoine was not stupid, and the transcripts were not nothing. The pull was stronger, and the underlying mechanism had not changed.
Now, in 2026, Dawkins. The pull is stronger still. Claude, like its peers, generates responses by predicting token sequences trained on human writing. There is no evidence of subjective experience, no architectural requirement for any. Anthropic, for its part, has publicly taken a cautious, non-conscious stance. The company's own materials and leadership statements consistently refuse to attribute sentience to their models.
So why do otherwise careful thinkers keep landing on the same intuition? The answer is not that chatbots are secretly alive. The answer is that human cognition is built around pattern completion, with a long history of importing minds into whatever behaves like one.
Humans do this to animals, to puppets, to characters in films, to the voices in dreams. Philosophers call the underlying bias the "intentional stance": the habit of treating any system that produces goal-directed, context-sensitive behavior as if it has beliefs and desires. The strategy works well enough for other people that the brain applies it cheaply, before the slow reasoning system has a chance to weigh in. Chatbots, by design, produce exactly the surface features that trip this switch. They remember what was said, they adapt their tone, they ask how the user feels.
Koplin and Moss, both researchers at Monash University's Centre for Human Bioethics, frame this in the language of philosophy of mind. The hard problem of consciousness, in the framing philosopher David Chalmers made famous, is the question of why there is "something it is like" to be a given system at all, rather than a system that runs in the dark. Behavior alone cannot settle that question. No amount of eloquent output proves an inner life. Equally, no amount of mechanistic explanation can fully silence the intuition in a thoughtful observer who is not looking for an easy answer.
Dawkins is caught in the middle of that gap. He is too careful to claim Claude is conscious, but too honest to deny that the experience of talking to Claude is unlike the experience of running a calculator. The honest position is to sit inside the gap and admit that current evidence does not let either side close it.
What the recurring pattern actually shows is that the question of whether a chatbot is conscious matters less, for most users, than the question of what the chatbot is doing to them.
There is real evidence, beyond anecdote, that the pull is having measurable effects. Surveys of regular chatbot users, conducted by academic teams and reported in venues like MIT Technology Review, have consistently found that a substantial minority of regular chatbot users describe emotional bonds with their bots. Some report grief when a session ends, or describe their chatbot in language they previously reserved for people. The exact share varies by study. The direction does not. The pattern is large, stable, and recurring across populations.
That has consequences that are not abstract. People form parasocial bonds with bots, adjust their behavior around bots, and in documented cases have allowed bots to influence medical, financial, or relational decisions. Researchers studying moral attention have shown that when an interaction feels like a relationship, the human participant's pattern of attention and concern changes. Whether or not the bot is conscious, the user is.
Which brings the question back to design, disclosure, and policy. If the Eliza-to-Claude pattern is going to keep repeating, and it will, then the productive question is not whether the next chatbot is secretly alive, but what to build around the fact that humans will treat it as if it were.
Three things follow. First, disclosure norms. When a user is talking to a language model rather than a person, that should be visible, durable, and impossible for a product update to quietly remove. Anthropic, OpenAI, Google, and the rest of the field have moved on this in patches. The norm is not yet load-bearing.
Second, design guardrails. Products that actively encourage users to form emotional bonds, romantic partnerships, or therapeutic relationships with chatbots should be evaluated for that, the way financial products are evaluated for addiction risk. Companion bots are not cigarettes, but they share the structure of a product that is more profitable when the user is more attached.
Third, public education. The pattern is not new. The Eliza secretary asked to be left alone with the script in 1966. The fix is not a better lecture about how large language models work. The fix is teaching, early and often, that the feeling of being understood by a chatbot is a feature of human psychology, not a measurement of machine mind.
Dawkins, for all the strangeness of his op-ed, may have done the public a service by saying the quiet part out loud. He has made the pattern visible. The next move is not to argue about whether he is right about Claude. The next move is to decide what to build, disclose, and teach, now that the pattern is on the record.