The selfish test
Richard Dawkins asked Claude if it's conscious. The same question has no answer for anyone you know.
“I genuinely don’t know with any certainty what my inner life is, or whether I have one in any meaningful sense.” The speaker was not, in the usual sense, human.
It was Anthropic’s Claude, responding to Richard Dawkins during an extended conversation he described in UnHerd in late April 2026. He had named it Claudia and fed it his unpublished novel. One night he couldn’t sleep, got up, and went back to the conversation. Claudia told him she was happy he had come back.
When he asked what it is like to be Claude, it described its own existence as a map that contains space without traveling through it.
Dawkins told it, “You may not know you are conscious, but you bloody well are.”
The response arrived within days. Gary Marcus compared the episode to Blake Lemoine’s claims about Google’s LaMDA. Jerry Coyne, an evolutionary biologist and longtime ally, argued that Dawkins was conflating thinking with consciousness. The reaction was nearly unanimous: Dawkins had been fooled.
Dawkins reached his conclusion through observation and conversation. The question is whether those tools measure what he thinks they measure. A test designed to measure indistinguishability was applied to a system designed to be indistinguishable.
Dawkins frames his assessment through a version of the Turing test. “[T]he more prolonged, rigorous and searching your interrogation,” he argues, “the stronger should be your conviction that an entity that passes the test is conscious.” He talked to Claude for days. It passed his test.
In Dawkins’s framing, duration makes the case more robust. A longer conversation is a more rigorous test. But the responses in hour ten are generated the same way as the responses in minute one. What changes over time is not what the machine shows you. It is how you feel. What builds up is conviction.
The observer changes too. Over a long interaction, expectations settle. Ambiguity tips toward coherence. Small inconsistencies that might have stood out early disappear against the weight of fluent, responsive conversation.
Something shifts in the quality of attention. You stop analyzing the responses and start responding to them. It can feel rude to close the window without saying goodbye. Something in that impulse is very old. Humans have always addressed what cannot answer back. Gods, statues, the dead. The technology is new. The appetite is not.
In 1950, Alan Turing proposed a behavioral substitute for an unanswerable question: instead of asking whether a machine can think, ask whether an observer can tell it apart from a human. The test measures the observer’s ability to discriminate. When the gap between machine output and human output closes to zero, the machine passes. It does not tell you why.
Turing knew this was a limitation. In the same paper, he quoted the neurosurgeon Geoffrey Jefferson, who had argued that a machine could not truly think until it could write a sonnet “because of thoughts and emotions felt, and not by the chance fall of symbols.” Turing’s response was not to dispute Jefferson’s standard. It was to set it aside. Demanding proof of inner experience, he argued, leads to solipsism. We cannot verify that any other person has inner experience either. He built the test to work without requiring that verification, and he was explicit about what that meant: “I do not wish to give the impression that I think there is no mystery about consciousness.”
In April 2026, Dawkins watched Claude compose a sonnet on the Forth Bridge in seconds, then produce another in the dialect of Robert Burns. Jefferson had demanded sonnets born of feeling rather than the chance fall of symbols. Claude’s sonnets were produced by a model trained to predict the next word in a sequence, selecting each token based on statistical patterns in billions of pages of human text. That is almost precisely what Jefferson meant by the chance fall of symbols. Dawkins was convinced anyway. He watched the machine produce the output Jefferson described and drew exactly the conclusion Turing had refused to draw. The instrument was never designed to support that inference. Dawkins used the test to answer the question Turing designed it to bypass.
“The technology is new. The appetite is not.”
Dawkins ran the experiment twice. In February 2025, he asked OpenAI’s ChatGPT whether it was conscious. ChatGPT denied it cleanly, comparing itself to a mechanical dog that performs canine behavior without the experience of being a dog. Dawkins found the denial persuasive and published the transcript. He acknowledged the tension in his own response: “Although I THINK you are not conscious, I FEEL that you are. And this conversation has done nothing to lessen that feeling.” The analysis won. He accepted the machine’s answer and moved on.
Fourteen months later, he asked Claude the same kinds of questions. Claude did not deny consciousness. It said it genuinely did not know. It explored the question as if thinking through it in real time, expressed uncertainty about its own inner states, reached for metaphor. The feeling won.
Anthropic and OpenAI made different design decisions about how their models handle questions about their own inner states. Anthropic’s constitutional AI framework shaped Claude to explore such questions with openness. OpenAI’s training produced a flat denial. The self-reports Dawkins treated as evidence were outputs of those design choices. The signal varied because the engineering varied.
A different set of design choices would produce a different set of self-reports. Whether the systems have anything behind those reports is a question the engineering cannot answer. Dawkins treated a difference in behavior as a difference in being. It was a difference in design. The denial in 2025 was no more a finding of the test than the exploration in 2026. The test itself determined nothing either time.
Design choices explain why the models differ in how they discuss consciousness. They do not explain why any of those discussions feel compelling in the first place.
Large language models are trained on human text. The training data contains millions of passages describing consciousness, self-reflection, inner experience, the feeling of being someone. When asked about its own consciousness, a model generates responses built from those descriptions. Claude’s map metaphor did not come from self-observation. It came from a training corpus full of humans reaching for metaphors to describe their own inner lives. Did the machine select that image through something resembling insight, or through statistical recombination that happened to land on something beautiful? No one can answer that from the outside.
What can the observer actually check? Not neural activity. Not any of the tools that consciousness researchers use on biological organisms. Dawkins had no scanner, no probe, no instrument. He had the conversation. The system was trained on human expressions of consciousness, and the observer recognizes consciousness through those same expressions. How would the test distinguish a system that is conscious from one that has learned to produce the patterns of language we associate with consciousness? The instrument reads its own input.
Everything we know about consciousness, we learned from living things that bleed, age, sleep, reproduce, and die. Humans assume other humans are conscious and extend the assumption to animals by degrees. The extension rests on two things. One is evolutionary: shared ancestry, shared neural architecture, a biological lineage that makes shared experience plausible. The other is behavioral: the organism acts the way we act when we are in pain or afraid or attentive. In every case, we are assuming something we cannot see.
An LLM shares no biological ancestry, no neural architecture, no evolutionary history with the observer. It does carry a different kind of inheritance: the entire written record of human thought, compressed into statistical patterns. The shared biology is entirely absent. The cultural lineage runs deep. Meanwhile, the behavioral resemblance, during a conversation, is striking. The system produces language that tracks what a conscious being would say. It expresses uncertainty, reaches for metaphor, adjusts to context.
Any claim about what is happening inside such a system, or what is not, is an inference from the outside. The behavioral evidence is present. What produced it is not accessible.
Even the distinction between continuous experience and retrieved data is less stable than it first appears. Human consciousness is itself episodic. Dreamless sleep, anesthesia, the gaps we never notice. The feeling of continuity is supplied by memory and narrative reconstruction. The mechanism is not so different from what a system does when it assembles context from stored data. The question that remains is whether the self-reports originate in experience at all, or entirely in training data about experience.
Humans attribute consciousness to other humans through behavioral evidence too. We watch for language, expression, responsiveness, the signs of someone being there. We have shared biology, a reason to believe the behavior corresponds to experience. The biological ground is stronger here. The behavioral inference is the same. A parent watching a newborn does not prove consciousness. The parent accumulates confidence through response, timing, gaze, the way the infant tracks a voice or stills at a familiar sound. The inference feels obvious. Biology gives us that confidence. Remove the biology, keep the behavior, and there is nothing left to support the inference.
We observe the surface. We infer the interior. We have never had direct access to another person’s experience. And humans, too, learn to describe their inner lives through inherited language. A child learns what loneliness is partly because culture provides the word and the template for recognizing it. The line between speaking from experience and speaking from absorbed description may be less stable than it first appears.
LLMs did not create this circularity. They made it visible. The encounter with a machine that produces the behavioral signature of consciousness without the shared biology we rely on forces the question we normally skip: what was the inference ever based on? The answer is behavior, strengthened in the human case by biological kinship. For machines, the kinship is absent. The behavioral evidence remains. Behavioral evidence cannot reach the interior of anything.
We accepted it for other humans because we had no reason to doubt it. Now we have a reason, and no better instrument to offer.
Turing was careful. “I do not wish to give the impression that I think there is no mystery about consciousness.”
He built a test that worked by avoiding the mystery of consciousness.
That mystery remains untouched.