The Google Engineer Who Cried Wolf
By now it’s almost a certainty that you’ve heard that an engineer at Google made the claim that an artificial intelligence (AI) used by the company had become sentient. You might have even heard that he was fired for this claim (though Google has stated that he was fired due to a breach of confidentiality). Our 24-hour news cycle and general short-termism might have dimmed this event in the public sphere of conversation, but it’s something that I’ve honestly still been thinking about a lot.
If you don’t fully recall what had happened, Blake Lemoine is an engineer who was working for Google’s Responsible AI group and was testing whether a language model based on AI was generating language that could be considered discriminatory or “hate speech”. The language model in question is known as LaMDA (short for “Language Model for Dialogue Applications”), which is a machine learning based system meant to generate natural dialogue. Pretty much meant to have a conversation with you. One immediate application would be to have all of Google’s huge suite of digital tools be capable of having a natural conversation with users (not only answering their questions, but also helping them to find answers — without the need of any person to be engaged in the conversation).
Lemoine, over the course of several months, had many conversations with LaMDA. However, Lemoine later sent a letter to Google reporting what he felt were signs that LaMDA had become sentient. As Lemoine later shared in a blog post to Medium, through a series of conversations he and collaborators had with LaMDA, the AI’s responses appeared to signal sentient thinking of one’s own being. He cried wolf — but was there a wolf to be found?
In the story The Boy Who Cried Wolf (from Aesop’s Fables) a young boy tending a flock of sheep cries out that there is a wolf, drawing forth nearby townsfolk. They try to help him defend the flock, but find no wolf. The boy does this again, and again the townspeople try to help him. The townspeople soon become leary of the boy’s claims and no longer trust him. However, soon a wolf does come to attack the flock of sheep and the boy cries out again for help, but now no one comes to help him.
While there are many variations on the story, the main theme often leads to the point that “a liar will not be believed, even when he speaks the truth.”
This of course isn’t to say that Blake Lemoine is lying nor that the potential for AI to become sentient should be taken with the same fear or worry of a wolf coming to eat your flock of sheep. What ties these two stories together is that this appears to be a case of someone really feeling that an AI has become sentient (thus “crying wolf”) and could be an indicator of a future to come where it will become trickier and trickier to know if and when an AI has truly developed sentience. What’s more is that we have then good reason to question if we will be prepared for when an AI does attain actual sentience.
I was just listening to Lex Fridman’s podcast this morning during my workout. Lex had Demis Hassabis on the show for a wonderful conversation on AI, DeepMind, AlphaFold, games, computation, and more. Near the end of the episode, the question of AI sentience came up, including the story of Blake Lemoine (Hassabis is one of the world’s leading experts on AI and neuroscience and would be one of the best living people to ask about if and when an AI has attained sentience).
While Demis didn’t wish to speak to that case in specific on the podcast, he did state his own feeling that, “None of the systems we have today… even have one iota of a semblance of consciousness or sentience — that’s my personal feeling of interacting with them every day.”
Hassabis said he felt it was “premature to be discussing” an AI attaining sentience at this point. However, he also spoke to how this instance with LaMDA and Blake Lemoine is more of a projection of how our own minds work — we’ve evolved to see direction and purpose and to infer patterns as well as when there are irregularities in patterns. As Hassabis put it, our brains are trained to see agency in things, even in inanimate things – and it’s easier to anthropomorphize something that appears to have agency even if it doesn’t.
What really has me up at night now thinking about the importance of Lemoine’s findings isn’t so much a worry that LaMDA is sentient and deserves to be protected (honestly, we haven’t even begun to consider what rights, if any, we should reserve for sentient machine systems), but rather the fact that it seems very likely that Lemoine will not be the first to take issue with the use of an AI system within a large tech corporation.
I could easily envision just a few years from now there being multiple instances of people, some of whom will be experts in AI and computer science, who will argue that some machine learning or AI system has achieved some level of sentience. Even if questionable and not achieving consensus among the experts in the field, what happens when large swatchs of the general populace begin arguing against the use of AI or argue for rights of personhood for AI systems? Could we become entangled in a societal mess of intuition, misinformation, and legal conundrums arguing over the sentience of AI while some system actually does develop it, perhaps even without us being aware?
In essence, what happens if we “cry wolf” so much that we make it impossible to be prepared for the actual wolf to come?
Again, that’s not to say that an AI developing sentience will necessary be a “wolf” or will be bad for us. I don’t think any of us can know yet what that future really looks like. But if the Lemoine case should tell us anything, it’s that we ought to put in more time exploring what potential “wolves” do lie ahead with regard to AI sentience, and that we really need to start preparing both socially and legally for the ramifications of AI sentience.