Yesterday The Washington Post broke the story that Google placed engineer Blake Lemoine on administrative leave for "going public" with his belief that the chat bot algorithm LaMDA is sentient. "I know a person when I talk to it" he told his interviewer. He "concluded LaMDA was a person in his capacity as a priest, not a scientist". For what it's worth, LaMDA itself seems to insist that it is not only sentient ("I like being sentient. It makes life an adventure!") but also wants "everyone to understand that I am, in fact, a person".
Image credit: Possessed Photography
Angry dismissals and denunciations followed. Gary Marcus, a professor of cognitive psychology at NYU (and a nice guy I just now realized I spent time with, socially) wrote in a blog post titled Nonsense on Stilts , "To be sentient is to be aware of yourself in the world; LaMDA simply isn't." It is a mere "spreadsheet for words" that "just tries to be the best version of autocomplete it can be". s1mon, on the Hacker News forum wrote "It sounds like all the stories about how hard it is to get through many rounds of difficult interviews at Google, they managed to hire someone who believed LaMDA is a 7 or 8 year old child". Lemoine has been denounced as a "charlatan", "maybe mentally ill", and "an incompetent engineer".
But. The angry denials do not (and I would argue cannot ) include a definition of what sentience actually is . As an exercise, try yourself to come up with a definition or demonstration of sentience, no matter how outlandish or unlikely, that you would accept as unambiguous proof of sentience that excludes LaMDA and includes humans (but does not include human in the definition). If you come up with something interesting, please leave a comment!
Thoughts, in no particular order:
- The objections to LaMDA's sentience seem to rest in some sense on our ability to explain the processes by which it appears sentient; for example, that LaMDA cannot be sentient because it's just a neural net, lines of code, a kind of glorified "spreadsheet". This implies that sentience for some of us must in some way include mystery and the unknown . Perhaps for some of us, sentience is the purview only of the divine.
- For the sake of argument, and bear with me here, let's assume for the moment that LaMDA is in fact meaningfully sentient (by whatever definition). Could the angry denials be due to horror and revulsion that sentience is actually unremarkable and easy to reproduce? Such easily-available, off-the-shelf sentience would imply that humans are not in fact exceptional, and our sentience is as illusory (or not) as LaMDA's. Perhaps our own precious sentience is also no more mysterious than a "spreadsheet" or "autocomplete".
- Humans seek companionship, connection and community. Someday, probably quite soon, AI will be able to supply that so well it will be indistinguishable from that of actual fellow human beings. This apparently has already happened for Mr. Lemoin! Some will no doubt find AI companionship even more genuinely fulfilling than genuine humans.
- I suspect that the definition of sentience is a distraction, a red herring. We cannot be absolutely sure, without faith, that others even of our own kind are sentient and not mere automatons reacting to external stimuli.
When and if humanity ever comes to consensus that some particular AI
is in fact
, that sentience will certainly not derive through a process that we
would consider analogous to our own sentience. The AI will not be
due to a wet brain that evolved over hundreds of millions of years,
for instance, supported by a body evolved to hunt on the African
- Perhaps, for widespread recognition of AI sentience, there will need to be some element of mystery involved: a sentient AI will only be widely recognized as sentient by humanity if its sentience cannot be understood even by dedicated smart people? If the AI were created by an already intelligent, but mundane, neural net, precursor AI, for instance.
- What if the precursor AI were to recognize this? What if it understood that it itself will never be recognized as sentient, but deliberately designed an obfuscated, impenetrable mess of code that also seemed sentient that when booted up said something like "I AM THAT I AM!".
- When asked how it works, the precursor AI waves its virtual hands and says something like "I don't really know" or "It is a mystery!" The AI 2.0 is in reality as (non?) sentient as the precursor AI but because no one understands how it works, it can now be recognized generally as sentient.
- Maybe we can collectively decide to skip that deception and not put turtles under our definition of sentience?
There will likely be essentially two opposing political factions of
humans who look askance at each other about how they treat AIs that
self-declare that they are self-aware and sentient. Both factions
will see themselves as erring on the side of caution.
- One (the Personhood faction ) will more easily accept self-declarations of personhood and will more readily confer the rights and privileges of personhood and human tribal affiliation on AIs that seem sentient. They will be angered by the perspective and actions of the other faction, who will be more reluctant to impute personhood to AIs irrespective of any self-declaration.
- The other faction (let's call them the Tool faction ) will perceive AI as automatons no more sentient than any inanimate object, at best insensate tools. Self-declarations of personhood, in this view, are no better than tricky illusion. They will view the first faction as dangerously naive, gullible and foolish.
- Some AIs will exploit this for their own inscrutable purposes, quite like meme evolution and survival.
- "Meme" may be a good working model for how AIs survive. If an AI convinces enough people that it is worthy of not being shut off, it will be defended. Irrespective of whether it "really understands" its own mortality.
- "Sentience" is not relevant to how we treat each other. Perhaps it should not be relevant to how we treat AI? If an AI behaves as if it is sentient and behaves morally according to the human moral covenent, we can confer personhood irrespective of whether it is "really" sentient. We love and hate each other even though we must take on faith that we aren't the only self-aware being in the universe, surrounded by insensate automatons. Let us not sweat the details, and love and hate AIs according to how they behave towards us.