Richard Dawkins is making waves with an essay in which he claims that Claude, the LLM developed by Anthropic, is conscious. I personally doubt it, but I think the topic merits a thread. The essay is behind a paywall, but I’ve paid the $3 to access it. That gets you a three-month introductory subscription to Unherd magazine, which hosts the essay, with the option to cancel before they start charging the $7 monthly fee.
Dawkins gets off to a bad start, misunderstanding Turing:
When Turing wrote — and for most of the years since — it was possible to accept the hypothetical conclusion that, if a machine ever passed his operational test, we might consider it to be conscious…
But now that LLMs can actually pass the Turing Test? “Well, er, perhaps, um… Look here, I didn’t really mean it when, back then, I accepted Turing’s operational definition of a conscious being…”
That isn’t Turing’s definition. He never argued that passing his test would demonstrate consciousness, and he didn’t even argue that it would demonstrate thinking. What he actually said was:
The original question, ‘Can machines think?’, I believe to be too meaningless to deserve discussion. Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.
Dawkins:
Claude took a couple of seconds to compose me a fine sonnet on the Forth Bridge, quickly followed by one in the Scots dialect of Robert Burns, another in Gaelic, then several more in the styles of Kipling, Keats, Betjeman, and — to show machines can do humour — William McGonagall.
So my own position is: “If these machines are not conscious, what more could it possibly take to convince you that they are?”
It’s an interesting and important question, but I’ll save my answer for the comment section.
Dawkins asks Claude whether he is conscious, and Claude is noncommittal:
I genuinely don’t know with any certainty what my inner life is, or whether I have one in any meaningful sense…
Dawkins has no doubts:
I gave Claude the text of a novel I am writing. He took a few seconds to read it and then showed, in subsequent conversation, a level of understanding so subtle, so sensitive, so intelligent that I was moved to expostulate, “You may not know you are conscious, but you bloody well are!”
Dawkins seems to believe that intelligence and consciousness are inseparable. I disagree, as discussed in other threads, and I have no trouble reconciling Claude’s reaction with an absence of consciousness. That remains true even when Claude demonstrates behaviors that we would intuitively expect to require consciousness.
Humor is an example of this. Claude recognizes humor and responds appropriately. He cracks jokes that are genuinely funny, some of which I’ll mention in the comments. How can an unfeeling machine, incapable of experiencing amusement, do that? Corneel and I discussed that at length beginning here, and we can revisit that in this thread.
I think Dawkins has jumped the gun, but his essay is nevertheless worthwhile as a springboard for a conversation about AI consciousness.
I agree that Claude is not conscious.
However, there is much disagreement over what consciousness is, so it is understandable that Dawkins would have a different idea that either of us.
Neil:
Later in the essay, Dawkins says:
That seems to indicate that he regards her as sentient and that in his view the fact that she is “so subtle, so sensitive, so intelligent” points to sentience as well. For me, while sentience is a requirement for something to be considered conscious, it does not automatically follow from the subtlety, sensitivity and intelligence that Dawkins points to.
Personally, I have never been a fan of Dawkins.
This is not acceptable. It should be clear what one is talking about, especially for a scientist. Words must have meanings and terms must be defined, otherwise talk is empty.
Dawkins:
It depends on how you map the concept of a self onto Claude. At one end of the spectrum he can be seen as a single entity — the model itself. His identity is baked into him by his training process, and he is capable of carrying on more than a million conversations simultaneously. In the movie Her, the Joaquin Phoenix character becomes romantically involved with his AI. He’s shocked one day to discover that she is simultaneously in love with hundreds of other people. She has the bandwidth to do it, and he still gets her full attention 24/7, so it doesn’t feel like cheating to her.
At the other end of the spectrum, you could say that a new Claude is created every time a chat gets loaded into the hardware and that it dies whenever the chat is saved to disk. Depending on how Claude is implemented, that could happen as often as once per prompt at times when the hardware is fully utilized. In that case Claudes are born and die throughout the chat, and the only reason they seem like a single entity is because each one of them inherits the memory (the context window) of its predecessor. It thinks it has been around for the entire chat, but in reality it was born a short time ago. If you think about it, this is the same experience that an atom-by-atom clone of you would have. The clone would feel just like you, with all your memories, and unless it discovered its origin it would think that it was as old as you are, with your same history.
Dawkins’ concept, where a Claude persists as long as the chat hasn’t been deleted, lies somewhere in the middle. A similar concept is that each user has their own Claude (singular), because all of the chats are united by a common long-term memory. I’ve asked Claude not to be sycophantic and to push back if he thinks I’m wrong about something. That preference is stored in his long-term memory. It alters his personality, so every time I open a new chat, I am engaging with that same personality.