June 2022. Blake Lemoine, an engineer at Google, claims that their new AI is sentient and is fired (The Verge).
Although, not quite. You can piece what actually happened from Lemoine’s own contemporary Medium article and the subsequent Washington Post piece [no paywall]: Lemoine shared a doc around Google titled “Is LaMDA Sentient?” (LaMDA is the name of the AI, a large language model like GPT-3) – a colleague said this was a bit provocative.
He started to speak with people outside the company and was placed on disciplinary leave for violating confidentiality. Lemoine upped the ante, inviting a lawyer to represent LaMDA,
and then you’re kinda done I reckon. But the point is that the question was asked.
Can an AI be sentient?
Are there already sentient AIs, and if not now then when? 1,000 years from now? Surely. 100 years? Probably. So 10 years? Maybe. How about 2025? Tomorrow?
How could we tell?
Would it matter?
I’m going to muddle sentience and consciousness here because I don’t want to get lost in definitions.
Wikipedia’s article on Sentience cites philosopher Antonio Damasio and says that sentience is a minimalistic way of defining consciousness
and limits it to the capacity to feel sensations and emotions.
According to this view: consciousness = sentience + creativity + intelligence + sapience + self-awareness + intentionality + more.
I’d prefer to say that our terms are ill-defined, and that consciousnesses may have all kinds of different characteristics, and may be a matter of degree.
So let’s enlarge the question, and agree to come back to pinning down terms later: can an AI be conscious?
2017. Philosopher Susan Schneider proposes ACT: the AI Consciousness Test.
The idea is that consciousness is something that is felt: we can all experience what it feels like, from the inside, to exist.
So the question for ACT is whether the synthetic minds we create have an experience-based understanding of the way it feels, from the inside, to be conscious.
i.e. do AIs feel the same as we do?
The proposed test is a series of questions.
Thus, the ACT would challenge an AI with a series of increasingly demanding natural language interactions to see how quickly and readily it can grasp and use concepts and scenarios based on the internal experiences we associate with consciousness. At the most elementary level we might simply ask the machine if it conceives of itself as anything other than its physical self. At a more advanced level, we might see how it deals with ideas and scenarios such as those mentioned in the previous paragraph. At an advanced level, its ability to reason about and discuss philosophical questions such as “the hard problem of consciousness” would be evaluated. At the most demanding level, we might see if the machine invents and uses such a consciousness-based concept on its own, without relying on human ideas and inputs.
(Article by Susan Schneider and Edwin Turner.)
One problem - as with GPT-3/ChatGPT - is that large language models are extraordinary mimics. So maybe they just say the right stuff to pass the test.
Schneider’s suggestion is to “box in” the AI away from human culture until we’ve tested it against the ACT, so it can’t make guesses.
I don’t know. I’m more convinced by the “quickly and readily” component of ACT. Surely there are some puzzles that are quicker to deduce if you have self-awareness? Dunno.
The AI Consciousness Test is one in a long line of tests for machine intelligence, such as the Turing Test.
2020. There’s a solid critique of ACT in this paper by David Udell and Eric Schwitzgebel, Susan Schneider’s Proposed Tests for AI Consciousness: Promising but Flawed (PDF at that link).
The challenge is that there’s always going to be a lower-level explanation of how the AI is answering questions on the silicon substrate (a giant lookup table, matrix maths, whatever), and that no series of questions is going to be sufficient to convince people that there is genuine machine consciousness at a higher level too.
One for the philosophers.
But Udell & Schitzgebel are articulate on the urgency of finessing ACT or something ACT-like:
AI consciousness, despite its present science-fictional air, may soon become an urgent practical issue. Within the next few decades, engineers might develop AI systems that some people, rightly or wrongly, claim have conscious experiences like ours. We will then face the question of whether such AI systems would deserve moral consideration akin to that we give to people. There is already an emerging ‘robot rights’ movement which would surely be energized by plausible claims of robot consciousness (Schwitzgebel and Garza 2015; Gunkel 2018; Ziesche and Yampolskiy 2019). So we need to think seriously in advance about how to test for consciousness among apparently conscious machines …
Schneider, in her Scientific American piece above, broadens the urgency to brain implants:
machine consciousness could impact the viability of brain-implant technologies, like those to be developed by Elon Musk’s new company, Neuralink. If AI cannot be conscious, then the parts of the brain responsible for consciousness could not be replaced with chips without causing a loss of consciousness. And, in a similar vein, a person couldn’t upload their brain to a computer to avoid death because that upload wouldn’t be a conscious being.
Consciousness is hard hey.
Consciousness is weird.
Let’s say that we agree that a silicon substrate can host consciousness.
Or that a group of organic cells, properly arranged etc, can host consciousness.
There is a slippery slope…
Eric Schwitzgebel again:
The United States is literally, like you, phenomenally conscious. That is, the United States literally possesses a stream of experiences over and above the experiences of its members considered individually.
If you’re a materialist, you probably think that rabbits have conscious experiences. And you ought to think that. After all, rabbits are a lot like us, biologically and neurophysiologically.
If you’re a materialist, you probably also think that conscious experience would be present in a wide range of naturally evolved alien beings behaviorally very similar to us even if they are physiologically very different. And you ought to think that. After all, it would be insupportable Earthly chauvinism to deny consciousness to alien species behaviorally very similar to us, even if they are physiologically different.
But, I will argue, a materialist who accepts consciousness in hypothetical weirdly formed aliens ought also to accept consciousness in spatially distributed group entities. If you then also accept rabbit consciousness, you ought also accept the possibility of consciousness in rather dumb group entities.
Finally, the United States is a rather dumb group entity of the relevant sort (or maybe even it’s rather smart, but that’s more than I need for my argument).
If we set aside our prejudices against spatially distributed group entities, we can see that the United States has all the types of properties that materialists normally regard as indicative of consciousness.
(I’ve added paragraph breaks.)
Schwitzgebel asks us to take the perspective of a consciousness entity which is much larger than us humans:
A planet-sized alien who squints might see the United States as a single, diffuse entity consuming bananas and automobiles, wiring up communication systems, touching the Moon, and regulating its smoggy exhalations – an entity that can be evaluated for the presence or absence of consciousness.
…and the rest of the chapter goes on to show convincingly that, yes, even if the USA isn’t conscious, it’s worthy of being evaluated.
(Do we need Schneider to write the USACT?)
This is perilously close to panpsychism, the view that the mind or a mindlike aspect is a fundamental and ubiquitous feature of reality.
We are conscious. My cat is conscious, although differently. Asteroids is conscious; AI is conscious, why not. Mud is conscious; a stellar nebula has its own nebula-like conscious. (Olaf Stapledon, in Star Maker, way back in 1937, wrote beautifully and poignantly about the culture of gas cloud megatheria at the dawn of the cosmos.)
What’s the alternative?
Maybe silicon can’t be conscious.
Maybe GPT-4, GPT-5, GPT-N, no matter how convincing, will be an AI p-zombie, a hypothetical being that is physically identical to and indistinguishable from a normal person but does not have conscious experience, qualia, or sentience.
Which implies there’s a cut-off somewhere. And I’m not happy with that either – I’m not ready to declare that my cat isn’t conscious, in her own cat way.
Everything is conscious.
Or nothing is conscious – except me. I’m not so sure about you.
Neither seems satisfying. Or useful?
Back to Eric Schwitzgebel, his paper (and forthcoming book) The Weirdness of the World, and the consciousness or otherwise of the USA…
Schwitzgebel asked philosopher Daniel Dennett, and he replied:
To the extent that the United States is radically unlike human beings, it’s unhelpful to ascribe consciousness to it. Its behavior is impoverished compared to ours and its functional architecture is radically unlike our own. Ascribing consciousness to the United States is not as much straightforwardly false is it is misleading. It invites the reader to too closely assimilate human architecture and group architecture.
And I like this approach, in a general sense, because it acknowledges the perspective from which we’re asking the question - being human - and therefore implicitly accepts that there will be other perspectives which have different answers.
The question is not: do we have conscious AIs?
It is more like: from our perspective, is there a non-misleading distinction between non-conscious AI and hypothetical conscious AI, and do we have conscious AIs in that sense?
AND THEN:
If an AI were to pass an AI Consciousness Test, in the non-misleading sense above, would it make any difference?
Udell & Schwitzgebel’s argument is that it’s meaningful in terms of robot rights.
But chickens have chicken-consciousness and we industrialise their growth and kill and eat them. Maybe the implication is that we ought to feel more gratitude when eating meat - if we eat meat at all - and that it’s poisonous to us to ignore that.
Or maybe they don’t have chicken-consciousness! Arguably we shouldn’t be treating chickens like we do in any case. It’s hard to imagine that we would treat them any worse even if we were certain they were lumps of 100% unthinking rock.
The point is that it’s not a question we really engage with, as a society. Maybe when it comes up with AI we collectively won’t care then, either.
So, for me, asking about AI consciousness is a way to winkle out these other questions.
Yes it’s important that we know when, in 50 years or 5 years, the machines wake up and we meet the first conscious AI. But if we then vary in our treatment of that AI, we’ll then have to ask what’s different about chickens, talking dogs, the Whanganui River in New Zealand which was granted legal personhood (BBC, 2017), the first uploaded nervous system - the open source OpenWorm virtual nemotode project - the entire USA as a conscious entity, and well, each other.
Definitely useful questions to ask.
June 2022. Blake Lemoine, an engineer at Google, claims that their new AI is sentient and is fired (The Verge).
Although, not quite. You can piece what actually happened from Lemoine’s own contemporary Medium article and the subsequent Washington Post piece [no paywall]: Lemoine shared a doc around Google titled “Is LaMDA Sentient?” (LaMDA is the name of the AI, a large language model like GPT-3) – a colleague said this was He started to speak with people outside the company and was placed on disciplinary leave for violating confidentiality. Lemoine upped the ante, and then you’re kinda done I reckon. But the point is that the question was asked.
Can an AI be sentient?
Are there already sentient AIs, and if not now then when? 1,000 years from now? Surely. 100 years? Probably. So 10 years? Maybe. How about 2025? Tomorrow?
How could we tell?
Would it matter?
I’m going to muddle sentience and consciousness here because I don’t want to get lost in definitions.
Wikipedia’s article on Sentience cites philosopher Antonio Damasio and says that and limits it to
According to this view: consciousness = sentience + creativity + intelligence + sapience + self-awareness + intentionality + more.
I’d prefer to say that our terms are ill-defined, and that consciousnesses may have all kinds of different characteristics, and may be a matter of degree.
So let’s enlarge the question, and agree to come back to pinning down terms later: can an AI be conscious?
2017. Philosopher Susan Schneider proposes ACT: the AI Consciousness Test.
The idea is that consciousness is something that is felt:
So the question for ACT is
i.e. do AIs feel the same as we do?
The proposed test is a series of questions.
(Article by Susan Schneider and Edwin Turner.)
One problem - as with GPT-3/ChatGPT - is that large language models are extraordinary mimics. So maybe they just say the right stuff to pass the test.
Schneider’s suggestion is to “box in” the AI away from human culture until we’ve tested it against the ACT, so it can’t make guesses.
I don’t know. I’m more convinced by the “quickly and readily” component of ACT. Surely there are some puzzles that are quicker to deduce if you have self-awareness? Dunno.
The AI Consciousness Test is one in a long line of tests for machine intelligence, such as the Turing Test.
2020. There’s a solid critique of ACT in this paper by David Udell and Eric Schwitzgebel, Susan Schneider’s Proposed Tests for AI Consciousness: Promising but Flawed (PDF at that link).
The challenge is that there’s always going to be a lower-level explanation of how the AI is answering questions on the silicon substrate (a giant lookup table, matrix maths, whatever), and that no series of questions is going to be sufficient to convince people that there is genuine machine consciousness at a higher level too.
One for the philosophers.
But Udell & Schitzgebel are articulate on the urgency of finessing ACT or something ACT-like:
Schneider, in her Scientific American piece above, broadens the urgency to brain implants:
Consciousness is hard hey.
Consciousness is weird.
Let’s say that we agree that a silicon substrate can host consciousness.
Or that a group of organic cells, properly arranged etc, can host consciousness.
There is a slippery slope…
Eric Schwitzgebel again:
(I’ve added paragraph breaks.)
Schwitzgebel asks us to take the perspective of a consciousness entity which is much larger than us humans:
…and the rest of the chapter goes on to show convincingly that, yes, even if the USA isn’t conscious, it’s worthy of being evaluated.
(Do we need Schneider to write the USACT?)
This is perilously close to panpsychism,
We are conscious. My cat is conscious, although differently. Asteroids is conscious; AI is conscious, why not. Mud is conscious; a stellar nebula has its own nebula-like conscious. (Olaf Stapledon, in Star Maker, way back in 1937, wrote beautifully and poignantly about the culture of gas cloud megatheria at the dawn of the cosmos.)
What’s the alternative?
Maybe silicon can’t be conscious.
Maybe GPT-4, GPT-5, GPT-N, no matter how convincing, will be an AI p-zombie,
Which implies there’s a cut-off somewhere. And I’m not happy with that either – I’m not ready to declare that my cat isn’t conscious, in her own cat way.
Everything is conscious.
Or nothing is conscious – except me. I’m not so sure about you.
Neither seems satisfying. Or useful?
Back to Eric Schwitzgebel, his paper (and forthcoming book) The Weirdness of the World, and the consciousness or otherwise of the USA…
Schwitzgebel asked philosopher Daniel Dennett, and he replied:
And I like this approach, in a general sense, because it acknowledges the perspective from which we’re asking the question - being human - and therefore implicitly accepts that there will be other perspectives which have different answers.
The question is not: do we have conscious AIs?
It is more like: from our perspective, is there a non-misleading distinction between non-conscious AI and hypothetical conscious AI, and do we have conscious AIs in that sense?
AND THEN:
If an AI were to pass an AI Consciousness Test, in the non-misleading sense above, would it make any difference?
Udell & Schwitzgebel’s argument is that it’s meaningful in terms of robot rights.
But chickens have chicken-consciousness and we industrialise their growth and kill and eat them. Maybe the implication is that we ought to feel more gratitude when eating meat - if we eat meat at all - and that it’s poisonous to us to ignore that.
Or maybe they don’t have chicken-consciousness! Arguably we shouldn’t be treating chickens like we do in any case. It’s hard to imagine that we would treat them any worse even if we were certain they were lumps of 100% unthinking rock.
The point is that it’s not a question we really engage with, as a society. Maybe when it comes up with AI we collectively won’t care then, either.
So, for me, asking about AI consciousness is a way to winkle out these other questions.
Yes it’s important that we know when, in 50 years or 5 years, the machines wake up and we meet the first conscious AI. But if we then vary in our treatment of that AI, we’ll then have to ask what’s different about chickens, talking dogs, the Whanganui River in New Zealand which was granted legal personhood (BBC, 2017), the first uploaded nervous system - the open source OpenWorm virtual nemotode project - the entire USA as a conscious entity, and well, each other.
Definitely useful questions to ask.