LaMDA is not sentient

You might have read news stories about a Google engineer who recently claimed that an AI language model (LaMDA) is sentient.

As many people have pointed out, LaMDA is definitely not sentient. They are correct. I also happen to think that the socio-cultural phenomenon of news articles about whether or not some certain AI is sentient, or conscious, or “human,” etc., makes for far more interesting research than an AI-sentience program would.

Nonetheless, this makes for a good opportunity to offer another take on why it is false to say that LaMDA is sentient. Typically, you’ll read responses (good ones!) — like those I linked above — which go at the question in (broadly) two kinds of ways (or both of those ways). The first way is to talk about how AI like LaMDA are mere “pattern-matchers”: they manipulate incredibly massive databases of language examples, which makes them really good autocompletes (as Gary Marcus points out here). The second way is to talk about how sentience depends on chemical signals (as Ned Block points out here). These are good, accurate points that I agree with. But the one I will make is completely different. It relies on Ludwig Wittgenstein.

I am sure helpful points could be made about what exactly we mean when we say that something is sentient. Let’s put this aside for now and assume our vague, varied intuitions are good enough (they are, for the point I’ll make). Now, a la Wittgenstein, let’s ask this: how are words like ‘sentient,’ ‘conscious,’ etc. used? How such words are used might tell us something about how comfortable we ought to be using such words with certain applications (this might or might not be a pun or both).

We can certainly make up stories about how “sentient” might have ended up being correctly used to talk about things like LaMDA.

Suppose some mysterious ancient being had created something like LaMDA — call it ALMDA (“ancient language model dialogue applications”) — some time before humans came around. Ignore all the superficial problems with that scenario (like how the mysterious ancient beings would have known English, for example), for a moment. Then suppose, at some point prior to our own computational inventions (or even theoretical formulations), we encountered ALMDA “in nature”. It probably matters, for this scenario, how we would have first encountered ALMDA (if it were in a human-lookalike vessel, for example, it is trivially reasonable to think we would have assumed ALMDA was sentient: we would have assumed it was human!). Let’s suppose, then, that we encountered some sort of console-with-instructions on how to communicate with ALMDA (or something such that we’d be unlikely to call it “human” or any other sentience-subsuming thing).

The takeaway I will suggest is this: we could well have referred to ALMDA — whatever we were talking with on the console (or perhaps the console itself, etc.) — as “sentient,” in such a scenario. We would, indeed, have had little reason to question ALMDA’s “sentience.” Even if we later found out exactly how ALMDA worked, nevertheless, the matter of sentience would be a non-question: ALMDA would have always been among the things which we’d always called sentient. Such “sentience” would simply be a part of our language.

LaMDA is not like ALMDA. We did not discover LaMDA in nature. We had a use for “sentient” prior to LaMDA, and prior to any like-AI. Similarly, our word “sentient” is not like the imagined “sentient” used in the ALMDA thought experiment. That word has an (imagined) entirely different historical evolution. It is a different word. To call LaMDA “sentient” is to attempt to make a different use out of our word “sentient.” However, it seems to me, nothing in our own linguistic history (up to this point) sanctions us to say that LaMDA (or any similar AI) is sentient. In other words, nothing sanctions this way of using the word “sentient”: Google engineers are not the final arbiters of meaning, and cannot gaslight signs into meaning whatever they wish. This is another reason to think it is false to say that LaMDA is sentient.

Posted by Steven McGannon