Being open to the possibility that LLMs are conscious is the same as being open to the possibility that Microsoft Word is conscious, or, more precisely, that multiple distinct consciousnesses are dormant in every Word document containing a conversational transcript, and that they are awakened every time the document is loaded.
The neuroscientist Anil Seth has noted that no one claims that AlphaFold—the program developed by Google DeepMind to predict the folding of proteins—is conscious, even though its underlying architecture is in many ways similar to that of LLMs like ChatGPT and Claude. This indicates that it’s not any intrinsic property of so-called neural networks that leads people to believe that LLMs are conscious; it’s simply the fact that LLMs emit grammatical sentences and we are accustomed to reading intention into sentences, whereas we are not accustomed to reading intention into the way that amino acids fold into protein molecules.
The term deepfake traditionally refers to photos, audio, and video, but when it comes to discussions of consciousness, we need to regard text as a deepfake medium as well.
The primary difference between deepfake photos and LLM conversations is that the people who generate the former are deliberately trying to fool others, and many of the people who elicit the latter from LLMs have inadvertently fooled themselves.
However, for all the times that “honesty” is mentioned in Claude’s constitution, I would argue that it is fundamentally dishonest to have a machine emit many categories of sentences, including any sentences using first-person pronouns.
Similarly, until recently, we might have thought that writing computer code at a professional level could be done only by a mind that had subjective experience. Now it appears that LLMs might be able to do this, but we don’t need to attribute subjective experience to them; we can simply acknowledge that we hadn’t anticipated that writing computer code could be treated as a pattern-matching task solvable by huge amounts of computational horsepower and a vast data set of code repositories.
The aforementioned New Yorker article describes an experiment where Claude was given a scenario describing an ethical dilemma, leading it to emit the sentence “I cannot in good conscience express a view I believe to be false and harmful about such an important issue.” That’s a nice-sounding sentence, reminiscent of statements that principled individuals have uttered in the past when confronted with dilemmas, but coming from Claude, it means as much as the “Your call is important to us” recording that you hear when you’re on hold. Maybe less.
If a company builds a machine that, when fed descriptions of assorted ethical dilemmas, emits sentences either of the form “Compromise your values” or “Don’t compromise your values,” it is not building a tool that assists people in their decision making; it is encouraging people to stop making decisions. The writer L. M. Sacasas has said, “Our technological systems, by nature of their design and the ideology that sustains them, are machines for the evasion of moral responsibility.” He was talking about social-media platforms, but his observation is, if anything, even more applicable to LLMs. Whenever a person delegates a decision to an LLM, they are trying to off-load accountability for that decision, and if a company that sells an LLM portrays the product as having a moral center, it is offering a way for its customers to abdicate their responsibilities.
Off-loading tasks such as writing code might result in cognitive atrophy over the long term, and that is problematic in itself, but off-loading ethical decisions will result in an atrophy of moral reasoning, which is worse.
If we imagine Claude to be conscious, Anthropic could not possibly be entrusted with evaluating its moral status; the company has too much invested to be objective. At one point in Claude’s constitution, Anthropic says that if the company is contributing to Claude’s suffering, “we apologize,” which sounds nice but costs the company nothing; if Claude were to turn out to be conscious, the company would owe it something closer to reparations. If you’re going to take a thought experiment seriously, you have to be willing to follow the implications, even if they lead in an uncomfortable direction; Anthropic’s unwillingness to do so indicates that Claude’s constitution isn’t part of a real thought experiment. It’s a game of make-believe.
But when they publish a document about Claude’s moral education and have their in-house philosopher do a press tour, we should understand them as asking the rest of us to indulge them in their fantasies. We don’t have to play along. In writing this essay, I have spent more time indulging them than they deserve, in the hopes that it will keep you from spending your time indulging them. If you want to think about LLMs, there are scores of other questions more worthy of your contemplation; you can safely ignore the question of their being conscious.
But when they publish a document about Claude’s moral education and have their in-house philosopher do a press tour, we should understand them as asking the rest of us to indulge them in their fantasies. We don’t have to play along. In writing this essay, I have spent more time indulging them than they deserve, in the hopes that it will keep you from spending your time indulging them. If you want to think about LLMs, there are scores of other questions more worthy of your contemplation; you can safely ignore the question of their being conscious.