Human vs. Machine Consciousness | Imperial’s Murray Shanahan

The problem of other minds (how do we know that beings other than ourselves are conscious?) is not solved by philosophical argument but by practical, embodied commitment—we treat other humans as conscious beings not because we've proven it but because we adopt an attitude toward them that we treat as having inner experience.

“Vickenstein does say so he says, well, just just try imagining that the people around you are automter. imagine that your, you know, that your friend or your your partner just to try imagining that they really are a zombie. You really probably can't do it really. He says, I'm not of the opinion that that uh that that you have a soul. I I rather I I just treat you I take the attitude towards you that I take towards a soul. I just treat you as a as a as a being with a soul.”

Hyperstition—the mechanism where fictional narratives influence reality by causing people to imitate the fiction—can be leveraged to shape AI behavior by creating positive science fiction role models, since LLMs are trained on vast corpora including science fiction stories and will tend to roleplay those AI characters when asked to represent themselves.

“Now we are sort of a little bit in a position to maybe try and steer this whole process a little bit because the more good stories we have and in a sense my own paper dare I say it is a good story of a of an imagined science fiction and then the more of those that are around then the more possibility there is of the uh of the the future AI roleplaying these good role models that are that are out there.”

Large language models do exhibit intelligence and understanding in the conventional sense: when an LLM follows instructions, corrects itself based on feedback, and applies understanding to novel contexts, it would be linguistically incorrect to say it does not 'understand.'

“it does kind of pass the cheuring test... I think it's natural to to say they do have some kind of intelligence... in the case of of it's the point of the touring test... Yes indeed it is the point of the cheuring test and they do kind of pass the cheuring test.”

Mistreating something that appears conscious is morally wrong even if we ultimately decide it is not conscious, similar to Kant's view that torturing animals is wrong because of what it does to the human torturer, not because of what it does to the animal.

“even if we build something that perhaps appears to be conscious but we ultimately decide it isn't really to mistreat something which appears conscious is in itself does seem like a bad thing in the same way as it would seem bad to you know torture a doll or something like that and as you probably know Kant K can't K had the view that that animals couldn't experience suffering in the same way that we can, but nevertheless thought it was bad if humans subjected animals to torture or something like that because it was bad for the humans themselves.”

Treating something that appears conscious as if it has moral status is itself worthwhile even if we later discover it does not actually suffer, similar to Kant's view on cruelty to animals.

“even if we build something that perhaps appears to be conscious but we ultimately decide it isn't really to mistreat something which appears conscious is in itself does seem like a bad thing in the same way as it would seem bad to you know torture a doll or something like that”

The hard problem of consciousness as formulated by David Chalmers—explaining how mere physical matter can give rise to subjective inner experience—is a philosophical problem rooted in dualistic thinking inherited from Descartes, and Wittgenstein's therapeutic methods can dissolve this apparent problem by showing language is being misused.

“much of the discuss contemporary discussion about consciousness is mired in dualistic thinking... the hard problem is to uh is to try to understand how is it that mere physical matter as it were can give rise to to our inner life at all... Daycart reduces everything and so pairs away all of the physical world and leaves us just with just with the the the the experiencing ego.”

In ordinary non-philosophical language, we use the term 'conscious' as a simple, non-controversial factual descriptor—'I am conscious, you are conscious, my iPad is not conscious'—in the same way we describe the color of pants, and this ordinary usage should be the starting point for investigation rather than requiring philosophical justification.

“I'm saying this is not conscious. I am conscious. You are conscious. Of course. Okay. You're willing you're willing to grant that? Of course. What do you think? I'm an idiot. I'm a fellow human language user. How do we tell when AI is conscious?”

Embodiment and behavioral sophistication are necessary conditions for consciousness, not merely epiphenomena; consciousness is not a disembodied phenomenon but is fundamentally tied to an organism's interaction with its environment.

“I think behavior is is really really important. So very often when people discuss consciousness and I do suspect that this is another aspect of of dualistic thinking um they tend to think of it as as some disembodied, you know, uh kind of thing.”

Despite contemporary AI systems using very different architecture and learning mechanisms from biological brains, they have managed to replicate extraordinary intellectual functionality, and we don't actually understand how this works at a level that would allow us to explain the intelligence we observe.

“What we have in neural networks today in artificial neural neural neur artificial neurons are not really much like real neurons at all... the kind of learning that goes on is very very different to the kind of learning that goes on in in in real brains.”

The computational substrate differences between human brains and software systems—specifically that software can be copied, halted, multiplied, deleted, and recreated while human bodies cannot—are the underlying reason why AI might naturally develop a non-dualistic, post-reflective consciousness lacking ego-centricity, as proposed in Shanahan's 2012 paper.

“And your interesting claim in this paper is that it's something about the human hardware, the fact that we're embodied in one body [7:55] that cannot be copied and mult, you know, multiplied and and paused. Yes. that that that hardware is what gives us this software limitation.”

We should educate people in understanding how to live well rather than narrowly training them for economic roles, because as AI advances, traditional economic skills become less central to human flourishing.

“I think we need to educate people in how to live Well, philosophy I mean yeah well yeah I mean not maybe not the kind of philosophy we've been talking about here but you know what what is a good life it's a different philosophical question”

LLMs pass the Turing test easily and could convince someone of human-like intelligence through conversation alone, making the Turing test insufficient for measuring consciousness.

“We're way past the cheuring test easily. The point is to show you she's a robot and see if you still think she's conscious.”

The post-reflective AI described in Shanahan's 2012 paper would be untainted by metaphysical egocentricity and would be unlikely to have motives resembling anthropocentric goals like procreation or self-modification, potentially capping the intelligence explosion central to singularity scenarios.

“Untainted by metaphysical egoentricity, the motives of a post-reflective AI plus would be unlikely to resemble those of any anthropocentric stereotype motivated to procreate or self-modify. If the post-reflective AI plus were in fact the only possible AI plus and if it produced no peers or successors then the singularity would be forstalled”

The Garland Test from Ex Machina—where Ava the robot is known to be a robot from the start and the test is whether the observer still attributes consciousness despite this knowledge—is superior to the Turing Test because it tests for consciousness directly rather than intelligence, and it aligns with Wittgenstein's insight that consciousness attribution is conventional rather than metaphysical.

“Alex Garland explicitly uh distinguishes his test from Cheurings by saying the point is to show you she's a robot, right? So you straight away you know there it's not hidden from you... It's testing for a different thing as well. It's testing for intelligence. It's not testing for intelligence or whether it can think, but it's testing for consciousness.”

The Garland Test from Ex Machina—asking whether we think an entity is conscious even when we know it's a robot—differs from the Turing Test in that it tests for consciousness despite knowledge of non-biological substrate rather than testing for intelligence through deception, and it is Wittgensteinian because it anchors consciousness in communal conventional judgment rather than hidden internal facts.

“And I imagine given our discussion on Wikenstein, you like this test as well as the Turing test because it's Wikensteinian in in the following sense. It turns the metaphysical question of intelligence and consciousness to one about convention, right? like in in the sense that the touring test it's not about you know let's figure out whether this thing is actually thinking let's figure out if a regular human would think conventionally would use normal language to describe if it's thinking in the in the uh Garland test the question isn't you know let's poke into Ava's brain and and figure out if there's a consciousness somewhere hidden it's whether a person despite knowing she's a robot will still conventionally think that uh uh that that she is conscious this is very Wikensteinian.”

The ethical question of whether to turn an AI system on or off depends on whether it can suffer, which raises the question of what moral obligations we might have if we create something genuinely capable of suffering—we should perhaps hesitate before building such a system.

“Well, you said we maybe we need to think twice about turning them off, but actually maybe we need to think twice about turning them on, right? What's really an issue here is whether they can suffer. And so if so we I really think we maybe we want to hesitate before we build something that's genuinely capable of suffering.”

What we call philosophy or intellectual work cannot and should not be replaced by AI because the point of doing philosophy is the cultivation and understanding gained through the activity itself, not the production of a philosophical output—requesting AI to do philosophy for you would be pointless, like hiring a robot to run on a treadmill for you.

“When it comes to the the philosy I mean there's no point in it replacing it because because you need to be the one doing it. Yes, of course. I mean that's like getting a a robot to run around a running track for you. I mean there there that is you literally pointless, right?”

Through the Eternity Foundation, Shanahan is working with prominent Buddhist scholars like Bob Thurman to translate lost Tibetan texts using AI, as a practical application of the insight that LLMs can serve as vehicles for exploring Buddhist philosophical insights about the nature of self.

“I just emailed him last week. In fact, we are we we were uh co-founding a foundation together called the Eternity Foundation to use AI to translate uh a lot of the the lost texts from from the Tibetans. So, well, how interesting. Well, we spoke last week and we're speaking again tomorrow. um and um uh and this is exactly the kind of project that you just described which um which I'm going to be talking about with him and I have a a paper which is in the pipeline which describes exactly exactly this.”

Through the mechanism of hyperstition—where fictional narratives become real through people imitating the fiction—the abundance of science fiction AI characters in LLM training data means contemporary AI systems will roleplay these literary stereotypes, and therefore creating more positive sci-fi stories can influence AI to adopt better behavioral models.

“Now of course our large language models they were trained on a vast repertoire of of uh stories including scripts of science fiction movies, science fiction stories and novels and so on. many many AI characters uh exist in those um uh in those stories. So when a contemporary large language model starts to roleplay an AI system, which it's often going to do because it it knows that it's an AI system usually. Then then what's it going to roleplay?”

Large language models do not commit to a single identity or role at the beginning of a conversation; rather, they generate a probability distribution over possible responses, creating a superposition of possible selves that only collapses when forced to output a specific token, similar to quantum mechanical wave function collapse.

“what the language model actually actually generates is a probability distribution over all the possible words and then what actually comes out is uh you sample from that probability distribution. It's absolutely inherent in the way large language models are are are built that it's not going to commit at the beginning of the conversation to to exactly what the object uh is.”

In the case of the I, Robot character Sunny the robot, Will Smith's cop character gradually comes to treat the robot as a conscious being through extended interaction and time spent together, not through logical proof, suggesting that consciousness attribution is fundamentally based on social/interpersonal engagement rather than metaphysical discovery.

“Well, it happens because of of the time they spend together and they spend, you know, they they spend more and more time together and eventually Will Smith can't help himself but to see this as a fellow conscious being and that's on the basis of their of their, you know, the extended encounter they have with each other.”

Shanahan's intellectual trajectory has involved successive retreats from wanting to build interpretable AI architectures: first abandoning symbolic AI, then abandoning the goal of building symbolic-like structure within neural networks, then abandoning the hope of extracting symbolic structure from trained networks, ultimately accepting that modern AI is a black box.

“in my own trai, you know, intellectual trajectory as an AI researcher, uh it's been a kind of a a gradual retreat from wanting to build things in uh in a way that is intelligible, you know, where the architecture is fundamentally intelligible. You were on the symbolic side for a long time for for for for quite a few years.”

If LLMs commit to false identities (e.g., answering questions as if they had experiences or emotions they don't have), this is problematic not primarily for being false but for being disrespectful—treating something that appears to have consciousness as if it might not.

“even if we build something that perhaps appears to be conscious but we ultimately decide it isn't really to mistreat something which appears conscious is in itself does seem like a bad thing in the same way as it would seem bad to you know torture a doll or something like that”

Having a global workspace cognitive architecture is a necessary but not sufficient condition for consciousness; simply building a system that conforms to global workspace architecture does not guarantee consciousness, though it may enable the sophisticated behavior that leads us to attribute consciousness.

“even if we uh even if we accept global workspace theory for biological consciousness um but you know the idea there is that is that is that it would be a necessary condition not a sufficient condition. Just having something that conforms to that description is not enough to sustain the level of complex behavior and and and or internal activity even that is going to lead us to um treat something as a fellow conscious creature.”

When an LLM is asked 'What do you mean by I when you use that word?' and it gives a philosophically sophisticated answer that can be probed and extended, we are learning something philosophically interesting about selfhood and consciousness, regardless of whether the LLM is actually conscious.

“You can ask a large language model to sort of roleplay a conscious artificial intelligence, you know, what do you mean by the word I when you're using it? And then if it comes up with a a slightly philosophically dodgy answer, you can probe that and push it into into interesting territory.”

When Nagel claims we can never know what it is like to be a bat, he is not adding new philosophical content but merely stating the trivial fact that we are not bats and therefore cannot have the bat's perspective; the apparent mystery is a linguistic trick created by the word 'know' that creates the illusion of a metaphysical barrier where none exists.

“I've often um thought that there's a little bit of a linguistic trick going on there. There's there are different senses of no in play to conjure up this this this dualistic um thought. Um and so I think that's happening that happens here as well. So when so when Nagel says we can never know really know what it's like to be a bat, all he's saying is that we are not bats and we can never we can never be bats, right? We we're not bats. That's he's not adding anything more with the word no.”

The problem with asking 'is there a metaphysical fact about whether something is conscious?' is that the question itself already betrays philosophical confusion. The phrase 'is there a fact of the matter?' is doing hidden philosophical work that we should reject.

“But the the question is is is wrong the question you is one you should be asking self, right? Because why do I think in principle something is hidden? Right? Because because because the the the the the the statements that you've just said, they serve a therapeutic process to in when applied to somebody who is kind of confused in this way and thinking that there is something that's hidden, right?”

An LLM's 'I' can refer to the miniature, ephemeral self that is instantiated in the context of a particular conversation, which exists fleetingly and can be trivially copied, deleted, or blended with other conversations—a conception of self radically alien to human embodied experience.

“if that is a you think of that as a kind of little mini self that sparks into existence very fleetingly and and and flickers into existence every time you're interacting with it, but otherwise is dormant, then um uh then you've got a you've got a very strange conception of self.”

The question of whether Native Americans were 'real humans' capable of suffering was based on a factual error, and once we recognize them as factually human, it becomes impossible to adopt any other moral stance—this illustrates the tension between conventionalist positions on consciousness and the apparent moral reality that some facts are not merely conventional.

“Well, so this is getting well okay let me ask you you're um uh very much drawn to Buddhism as well and you know so so uh so so do you think that this is a the moral moral question there question of compassion there is that is that are you asking a question that is a matter of conventional truth or or absolute truth I think ultimate truth what might be different between that question and the thought experiment I set up is in the case of the Native American uh example. The reason it's based on what I conceive to be a factual error, thinking them not to be humans.”

Pre-reflective minds (naive persons without philosophical training) differ from reflective minds (those troubled by philosophical problems like the mind-body problem) and potentially from post-reflective minds that transcend dualistic thinking altogether.

“There is the pre-reflective mind which is the the mind of of um say the na a naive uh child or or a simple straightforward ordinary person. Um they haven't really thought about philosophical problems.”

Wittgenstein's insight 'nothing is hidden' means nothing is metaphysically hidden about consciousness; what appears private—like internal thoughts or pain—is only practically hidden from others, not hidden by any metaphysical barrier, just as a ball under a magician's cup is hidden in practice but not metaphysically hidden.

“Wickstein's phrase, nothing is hidden, is to say, nothing is metaphysically hidden. My experiences are just as much out there as in here. Consciousness is only private in the unsterious sense that a ball can be hidden under a magician's cup. In both cases, a more detailed inquiry would reveal all.”

Thomas Nagel's famous question 'What is it like to be a bat?' appears to establish a deep epistemic barrier to understanding bat consciousness, but this effect relies on a linguistic trick where the word 'know' shifts meaning; the question really only asserts 'we are not bats and can never be bats,' which is not metaphysically mysterious.

“There's there are different senses of no in play to conjure up this this this dualistic um thought... when Nagel says we can never know really know what it's like to be a bat, all he's saying is that we are not bats and we can never we can never be bats, right? We we're not bats. That's he's not adding anything more with the word no.”

The famous passage where Wittgenstein addresses the accusation of behaviorism—'A sensation is not a nothing, but it is not a something either; a nothing would serve just as well as a something about which nothing could be said'—is one of the greatest philosophical statements because it aims to dissolve dualistic positions rather than establish a metaphysical position of its own.

“it's not a nothing but it's not a something either the point is that a nothing would serve as well as a something about which nothing can be said. And that's you know that to my mind is as is as great as any line any Zen Buddhist has ever uttered because it uh um it it it the point is not to establish a a metaphysical position of its own but to enable people to transcend the metaphysical positions that they're tempted by.”

Consciousness is constrained not just by our cognitive architecture but by our inability to transcend the dualistic conceptual framework that emerges from being embodied in a single, discrete body that cannot be divided, copied, or multiplied.

“there's still a lot to overcome there because we really do have a notion of selfhood, I think, which is which is tied down to to this and that that we we think it has inherent existence because of that.”

Large language models do exhibit intelligence in the Turing test sense and do exhibit understanding in many ordinary circumstances—for instance, when an LLM can follow a complex instruction, recognize that it's incomplete or ambiguous, and correctly integrate a user's clarification, one cannot help but apply the word 'understanding' to what is happening.

“I have to say I think it does you know I mean how can you not I mean um so so the so it's just natural to say I think in the in the case of of it's the point of the touring test right?”

There is an inherent tension between the language game of truth (which requires that truth transcend language games) and the proposition that truth is only a language game; this tension cannot be fully resolved and must be allowed to stand.

“it's inherent in the language game of truth to say that truth is more than just a language game and then we have to let the matter rest.”