Interview on whether machines can be conscious with Scott Aaronson, theoretical computer scientist and David J. Bruton Jr. Centennial Professor of Computer Science at the University of Texas at Austin. His primary areas of research are quantum computing and computational complexity theory.
Scott blogged about this and other segments of our interview – his blog is very popular and has way more comments than this one does – check it out.

Transcript

Adam Ford: In ‘Could a Quantum Computer have Subjective Experience?‘ you speculate where the process has to fully participate in the arrow of time to be conscious and this points to decoherence. If pressed, how might you try to formalize this?

Scott Aaronson: So yeah so I did write this kind of crazy essay five or six years ago that was called “The Ghost in the Quantum Turing Machine“, where I tried to explore a position that seemed to me to be mysteriously under-explored! And all of the debates about ‘could a machine be conscious?’ and we want to be thoroughgoing materialists right? There’s no magical ghost that defies the laws of physics; the brains or physical systems that obey the laws physics
just like any others.
But there is at least one very interesting difference between a brain and any digital computer that’s ever been built – and that is that the state of a brain is not obviously copyable; that is not obviously knowable to an outside person well enough to predict what a person will do in the future, without having to scan the person’s brain so invasively that you would kill them okay. And so there is a sort of privacy or opacity if you like to a brain that there is not to a piece of code running on a digital computer.
And so there are all sorts of classic philosophical conundrums that play on that difference. For example suppose that a human-level AI does eventually become possible and we have simulated people who were running a inside of our computers – well if I were to murder such a person in the sense of deleting their file is that okay as long as I kept the backup somewhere? As long as I can just restore them from backup? Or what if I’m running two exact copies of the program on two computers next to each other – is that instantiating two consciousnesses? Or is it really just one consciousness? Because there’s nothing to distinguish the one from the other?
So could I blackmail an AI to do what I wanted by saying even if I don’t have access to you as an AI, I’m gonna say if you don’t give me a million dollars then I’m just going to – since I have your code – I’m gonna create a million copies of your of the code and torture them? And – if you think about it – you are almost certain to be one of those copies because there’s far more of them than there are of you, and they’re all identical!
So yeah so there’s all these puzzles that philosophers have wondered about for generations about: the nature of identity, how does identity persist across time, can it be duplicated across space, and somehow in a world with copy-able AIs they would all become much more real!
And so one one point of view that you could take is that: well if I can predict exactly what someone is going to do right – and I don’t mean you know just saying as a philosophical matter that I could predict your actions if I were a Laplace demon and I knew the complete state of the universe right, because I don’t in fact know the complete state of the universe okay – but imagine that I could do that as an actual practical matter – I could build an actual machine that would perfectly predict down to the last detail every thing you would do before you had done it.
Okay well then in what sense do I still have to respect your personhood? I mean I could just say I have unmasked you as a machine; I mean my simulation has every bit as much right to personhood as you do at this point right – or maybe they’re just two different instantiations of the same thing.
So another possibility, you could say, is that maybe what we like to think of is consciousness only resides in those physical systems that for whatever reason are uncopyable – that if you try to make a perfect copy then you know you would ultimately run into what we call the no-cloning theorem in quantum mechanics that says that: you cannot copy the exact physical state of a an unknown system for quantum mechanical reasons. And so this would suggest of you where kind of personal identity is very much bound up with the flow of time; with things that happen that are evanescent; that can never happen again exactly the same way because the world will never reach exactly the same configuration.
A related puzzle concerns well: what if I took your conscious or took an AI and I ran it on a reversible computer? Now some people believe that any appropriate simulation brings about consciousness – which is a position that you can take. But now what if I ran the simulation backwards – as I can always do on a reversible computer? What if I ran the simulation, I computed it and then I uncomputed it? Now have I caused nothing to have happened? Or did I cause one forward consciousness, and then one backward consciousness – whatever that means? Did it have a different character from the forward consciousness?
But we know a whole class of phenomena that in practice can only ever happen in one direction in time – and these are thermodynamic phenomena right; these are phenomena that create waste heat; create entropy; that may take these little small microscopic unknowable degrees of freedom and then amplify them to macroscopic scale. And in principle there was macroscopic records could could get could become microscopic again. Like if I make a measurement of a quantum state at least according to the let’s say many-worlds quantum mechanics in principle that measurement could always be undone. And yet in practice we never see those things happen – for the same for basically the same reasons why we never see an egg spontaneously unscramble itself, or why we why we never see a shattered glass leap up to the table and reassemble itself right, namely these would represent vastly improbable decreases of entropy okay. And so the speculation was that maybe this sort of irreversibility in this increase of entropy that we see in all the ordinary physical processes and in particular in our own brains, maybe that’s important to consciousness?
Right uh or what we like to think of as free will – I mean we certainly don’t have an example to say that it isn’t – but you know the truth of the matter is I don’t know I mean I set out all the thoughts that I had about it in this essay five years ago and then having written it I decided that I had enough of metaphysics, it made my head hurt too much, and I was going to go back to the better defined questions in math and science.

Adam Ford: In ‘Is Information Physical?’ you note that if a system crosses a Swartzschild Bound it collapses into a black-hole – do you think this could be used to put an upper-bound on the amount of consciousness in any given physical system?

Scott Aaronson: Well so I can decompose your question a little bit. So there is what quantum gravity considerations let you do, it is believed today, is put a universal bound on how much computation can be going on in a physical system of a given size, and also how many bits can be stored there. And I the bounds are precise enough that I can just tell you what they are. So it appears that a physical system you know, that’s let’s say surrounded by a sphere of a given surface area, can store at most about 10 to the 69 bits, or rather 10 to the 69 qubits per square meter of surface area of the enclosing boundary. And it has a similar limit on how many computational steps it can do over it’s it’s whole history.
So now I think your question kind of reduces to the question: Can we upper-bound how much consciousness there is in a physical system – whatever that means – in terms of how much computation is going on in it; or in terms of how many bits are there? And that’s a little hard for me to think about because I don’t know what we mean by amount of consciousness right? Like am I ten times more conscious than a frog? Am I a hundred times more conscious? I don’t know – I mean some of the time I feel less conscious than a frog right.
But I am sympathetic to the idea that: there is some minimum of computational interestingness in any system that we would like to talk about as being conscious. So there is this ancient speculation of panpsychism, that would say that every electron, every atom is conscious – and do me that’s fine – you can speculate that if you want. We know nothing to rule it out; there were no physical laws attached to consciousness that would tell us that it’s impossible. The question is just what does it buy you to suppose that? What does it explain? And in the case of the electron I’m not sure that it explains anything!
Now you could say does it even explain anything to suppose that we’re conscious? But and maybe at least not for anyone beyond ourselves. You could say there’s this ancient conundrum that we each know that we’re conscious presumably by our own subjective experience and as far as we know everyone else might be an automaton – which if you really think about that consistently it could lead you to become a solipsist. So Allen Turing in his famous 1950 paper that proposed the Turing test had this wonderful remark about it – which was something like – ‘A’ is liable to think that ‘A’ thinks while ‘B’ does not, while ‘B’ is liable to think ‘B’ thinks but ‘A’ does not. But in practice it is customary to adopt the polite convention that everyone thinks. So it was a very British way of putting it to me right. We adopt the polite convention that solipsism is false; that people who can, or any entities let’s say, that can exhibit complex behaviors or goal-directed intelligent behaviors that are like ours are probably conscious like we are. And that’s a criterion that would apply to other people it would not apply to electrons (I don’t think), and it’s plausible that there is some bare minimum of computation in any entity to which that criterion would apply.

Adam Ford: Sabine Hossenfelder – I forget her name now – {Sabine Hossenfelder yes} – she had a scathing review of panpsychism recently, did you read that?

Scott Aaronson: If it was very recent then I probably didn’t read it – I mean I did read an excerpt where she was saying that like Panpsychism – is what she’s saying that it’s experimentally ruled out? If she was saying that I don’t agree with that – know I don’t even see how you would experimentally rule out such a thing; I mean you’re free to postulate as much consciousness as you want on the head of a pin – I would just say well it’s not if it doesn’t have
an empirical consequence; if it’s not affecting the world; if it’s not affecting the behavior of that head of a pin, in a way that you can detect – then Occam’s razor just itches to slice it out from our description of the world – always that’s the way that I would put it personally.\
So I put a detailed critique of integrated information theory (IIT), which is Giulio Tononi’s proposed theory of consciousness on my blog, and my critique was basically: so Tononi know comes up with a specific numerical measure that he calls ‘Phi’ and he claims that a system should be regarded as conscious if and only if the Phi is large. Now the actual definition of Phi has changed over time – it’s changed from one paper to another, and it’s not always clear how to apply it and there are many technical objections that could be raised against this criterion. But you know what I respect about IIT is that at least it sticks its neck out right. It proposes this very clear criterion, you know are we always much clearer than competing accounts do right – to tell you this is which physical systems you should regard as conscious and which not.
Now the danger of sticking your neck out is that it can get cut off right – and indeed I think that IIT is not only falsifiable but falsified, because as soon as this criterion is written down (what the point I was making is that) it is easy to construct physical systems that have enormous values of Phi – much much larger then a human has – that I don’t think anyone would really want to regard as intelligent let alone conscious or even very interesting.
And so my examples show that basically Phi is large if and only if your system has a lot of interconnection – if it’s very hard to decompose into two components that interact with each other only weakly – and so you have a high degree of information integration. And so my the point of my counter examples was to try to say well this cannot possibly be the sole relevant criterion, because a standard error correcting code as is used for example on every compact disc also has an enormous amount of information integration – but should we therefore say that you know ‘every error correcting code that gets implemented in some piece of electronics is conscious?’, and even more than that like a giant grid of logic gates just sitting there doing nothing would have a very large value of Phi – and we can multiply examples like that.
And so Tononi then posted a big response to my critique and his response was basically: well you’re just relying on intuition; you’re just saying oh well yeah these systems are not a conscious because my intuition says that they aren’t – but .. that’s parochial right – why should you expect a theory of consciousness to accord with your intuition and he just then just went ahead and said yes the error correcting code is consciouss, yes the giant grid of XOR gates is conscious – and if they have a thousand times larger value of Phi than a brain, then there are a thousand times more conscious than a human is. So you know the way I described it was he didn’t just bite the bullet he just devoured like a bullet sandwich with mustard. Which was not what I was expecting but now the critique that I’m saying that ‘any scientific theory has to accord with intuition’ – I think that is completely mistaken; I think that’s really a mischaracterization of what I think right.
I mean I’ll be the very first to tell you that science has overturned common sense intuition over and over and over right. I mean like for example temperature feels like an intrinsic quality of a of a material; it doesn’t feel like it has anything to do with motion with the atoms jiggling around at a certain speed – okay but we now know that it does. But when scientists first arrived at that modern conception of temperature in the eighteen hundreds, what was essential was that at least you know that new criterion agreed with the old criterion that fire is hotter than ice right – so at least in the cases where we knew what we meant by hot or cold – the new definition agreed with the old definition. And then the new definition went further to tell us many counterintuitive things that we didn’t know before right – but at least that it reproduced the way in which we were using words previously okay.
Even when Copernicus and Galileo where he discovered that the earth is orbiting the Sun right, the new theory was able to account for our observation that we were not flying off the earth – it said that’s exactly what you would expect to have happened even in the in ?Anakin? because of these new principles of inertia and so on okay.
But if a theory of consciousness says that this giant blank wall or this grid is highly highly conscious just sitting there doing nothing – whereas even a simulated person or an AI that passes the Turing test would not be conscious if it’s organized in such a way that it happens to have a low value of Phi – I say okay the burden is on you to prove to me that this Phi notion that you have defined has anything whatsoever to do with what I was calling consciousness you haven’t even shown me any cases where they agree with each other where I should therefore extrapolate to the hard cases; the ones where I lack an intuition – like at what point is an embryo conscious? or when is an AI conscious? I mean it’s like the theory seems to have gotten wrong the only things that it could have possibly gotten right, and so then at that point I think there is nothing to compel a skeptic to say that this particular quantity Phi has anything to do with consciousness.