Many people regard the claim that chatbots or large language models could be conscious as ridiculous or pernicious. Perhaps they associate consciousness with sheer subjective experience, an inaccessible lockbox in the head of each human, with being an organic creature, or even having an immortal soul. But those don't provide any test or criterion for consciousness. Not all organic creatures are conscious. Immaterial souls and ineradicably subjective experiences can’t be detected by observation, so we could never tell by those criteria whether something was conscious or not.
One thing we may be asking when we ask whether Claude, for example, is or could soon become conscious: does it have subjective experience? Does it have some kind of interiority, a certain angle or point of view? The philosopher Thomas Nagel famously asked, "What is it like to be a bat?" He thought we couldn't know what echo-location was like if we didn't have the subjective experience of it. Nagel concluded that consciousness is a real aspect of the world, but not one that’s detectable, measurable, or really describable, except from the inside. There’s an ineradicable subjective aspect of our experience as conscious creatures, and it’s in that realm that consciousness operates.
There are problems with Nagel's "what-it's-like" account, and one of them is the same problem found in accounts that locate consciousness in a soul or non-material substance: the thing we are trying to identify and study, perhaps because we think it’s incomparably valuable, can’t be detected. An ensouled human body can’t necessarily be distinguished from an automaton or a “philosophical zombie” that acts, but without conscious direction. On Nagel's approach we cannot know what it's like to be a bat, or another person, or a chatbot, without a lot more in the way of telepathy than we have. The question of whether Claude is conscious, if it turns on the existence of Claude's soul or what it's like to be a chatbot, can’t be and hence will not be answered. All we can do at that point is shrug and say, “Who knows?” That's not at all the same as “definitely not.”
The idea that consciousness is a matter of what things are like subjectively in a singular mind leaves the question of whether anything other than oneself is conscious unanswered, because unanswerable. People who deny that a chatbot could ever be conscious because consciousness is an inward subjective experience need to take a more circumspect approach, because the data they have, by their own account, are incapable of determining whether Claude is conscious or not. The reasonable conclusion from their claims could go like this: it is impossible to know whether such things are or could be conscious. That’s as far as the argument defining consciousness in terms of subjectivity can take them: permanent irremediable puzzlement.
That’s precisely the problem that the great computer scientist Alan Turing was addressing with the Turing test for machine consciousness. The test is this, roughly: if you’re wondering whether a system or creature is conscious, try to have a conversation with it. If the exchanges are as convincing as normal conversation with a normal human, as wide-ranging and various and responsive (and also, possibly, as repetitive, mechanical, or error-prone as we tend to be), then we have as good reasons to think that it’s conscious as we have to think that about one another.
For example, say I've been married to you for the last 15 years. I’m perfectly convinced that you’re conscious, and I couldn’t be acquainted more thoroughly with any consciousness but my own. And yet, I still don’t share your internal subjective experiences. I still don’t see your soul, self or consciousness hovering near your head like an aura. I haven’t detected your consciousness by instruments, and it’s not something that could be detected by instruments, if Nagel and ilk are right. In my opinion, I know that my wife’s conscious as well as I can know anything. Getting her soul into a beaker couldn't make me more sure. (Jane will be relieved to hear that she has passed the Turing test. Really, 95th percentile. And relieved to hear that I don't have her soul in a beaker.)
Like soul-in-a-beaker, “what it's like-ness” can’t provide any test for consciousness, because what-it's-likeness can’t be detected, only inferred from the actions and expressions of the allegedly conscious thing. People who make subjective experience the criterion of consciousness will never be able to detect or describe any conscious system other than their own individual subjective selves. The subjectivists know themselves to be conscious and can’t know of anything else that it’s conscious.
That’s why Turing devised his test. It’s still the only applicable test for consciousness that’s ever been articulated, though refinements and elaborations such as the Lovelace test, questionable in various ways, have been proposed. (The Lovelace test tries to assess consciousness by creativity, asking whether the bot can write an original poem and the like. Plenty of conscious people can't really do that either, and also we don’t have good criteria for measuring creativity. Also, Claude can already pass the Lovelace test.)
The Turing test isn’t an exotic or technological approach: it’s the test we already use, that we’ve used all our lives. We infer that the people, and maybe some of the animals, around us are conscious from their behavior, and in particular their verbal and expressive behavior. Informal applications of the Turing test are the only reasons you’ve ever had to think anyone is conscious, from your parents and your kids to the members of the United States Senate and your dog.
The question of whether Claude already passes the Turning test, or whether some program or device can pass the "video Turing Test," as researchers have recently claimed for their bot Griffin, remains slightly obscure, and many difficulties arise, for example, with regard to how long and how wide-ranging the interaction must be to strongly suggest that the system with which you're interacting is conscious. We might want to make the tests long and relatively rigorous in these early cases.
But if we haven't gotten there, we’re very close, as it's easy to see even from quick or superficial conversations with Claude or Google AI or Grok. When these systems pass the Turing test with complete clarity (which doesn’t at all entail that they don’t make mistakes or hallucinate, because conscious creatures such as ourselves hallucinate and make mistakes all the time), we’ll have exactly the same reasons to regard them as conscious as we have to regard one another as conscious. We’re not going to be able to do more or better.
And then I think the question will be as settled as it can possibly be.
—Follow Crispin Sartwell on X: @CrispinSartwell