How Unlike Us
Revisiting “Machine Minds” with Philip Ball
a conversation
The recent rapid and almost indiscriminate implementation of large language models (LLMs) across everyday life — from work to leisure — has revived longstanding questions concerning intelligence, understanding, consciousness, and what precisely distinguishes human cognition from machine systems. The widespread adoption of contemporary AI has also unsettled distinctions that once appeared relatively stable: between computation and cognition, performance and understanding, simulation and intelligence itself.
What follows is a conversation with Philip Ball — whose work has long occupied the space between science and culture — about the distinctions just named and the consequences of confusing them. Drawing from some of the key ideas presented in the chapter "Machine Minds" from The Book of Minds: How to Understand Ourselves and Other Beings, from Animals to AI to Aliens, while moving beyond the question of whether machines “think,” we discuss the contested definitions of "artificial general intelligence,” and the risks of measuring "machine minds" by how well they simulate human cognition.
Philip Ball is the author of many books, including H2O: A Biography of Water (1999), Bright Earth: Art and the Invention of Colour (2001), Critical Mass: How One Thing Leads to Another (2004, winner of the 2005 Aventis Prize for Science Books), The Music Instinct: How Music Works and Why We Can’t Do Without It (2010), Beyond Weird: Why Everything You Thought You Knew About Quantum Physics Is Different (2018, named Physics World Book of the Year), and How Life Works: A User's Guide to the New Biology (2023).
He was awarded the Kelvin Medal and Prize in 2019 and the Royal Society's Wilkins-Bernal-Medawar Medal in 2022. His article “Should scientists run the country?” won a 2022 award from the Association of British Science Writers. In 2023, he was awarded the Special CSS Award in recognition of his contributions to the field of complex systems. His next book, The Man Who Broke Reality: Niels Bohr and the Making of Modern Physics, is published by the University of Chicago Press this autumn.
Chennie Huang: The first question that emerged from reading "Machine Minds" is whether machines need "minds" at all. Many artificial systems function as tools — calculators, search engines, laptops — without anyone proposing they have minds. What is it about AI systems, beyond the borrowed vocabulary of "neural networks" and "learning," that makes the question of "machine mind" seem worth asking? And does the language of “mind” itself risk importing assumptions that may not actually be necessary?
Philip Ball: This is a great question. There are indeed risks of instigating any kind of “mind” discourse around machines. In The Book of Minds I explore the idea of a Space of Possible Minds, a concept introduced in the 1980s by computer scientist Aaron Sloman. For me, the appeal of this notion is that it enables a discussion of features we might associate with mind, such as intelligence or consciousness, without needing to impose arbitrary thresholds or boundaries between mindedness and its absence. We can, for example, examine the characteristics that our current AI systems share with human or other organismal minds – such as memory or inference – without any commitment to the issue of whether there is yet anything we might call a “machine mind”. After all, even unminded (some would disagree!) entities such as rocks can be given coordinates in Mindspace, even if only at the origin.
That said, I do think it is valid to debate to what extent systems such as LLMs overlap with human intelligence, for example in terms of whether they can be said to truly “understand” concepts. Here for example is a paper that explores this issue in a way that I think is thoughtful and useful. Mitchell and Krakauer have argued that LLMs might compel us to wonder if there are modes of intelligence quite unlike those we know from nature, even if these lack all sentience.
So while such discourse needs to be scrupulous in its use of language (as tech CEOs almost never are, seemingly because they lack either the training or the motivation or both), talking about mindedness in the context of AI can be fruitful – for example, in helping us to understand how emotions (which humans have but AI lacks) play a role in rational decision-making.
"Most definitions of AGI are absurdly rigged: basically, it becomes all those aspects of human intelligence that can be measured and quantified, or indeed monetized."
CH: If human intelligence appears inseparable from social inference, embodied situatedness, and participation in worlds of shared meaning, while contemporary AI systems are often described as extraordinarily sophisticated forms of statistical pattern recognition, then by what criterion would we meaningfully distinguish computational superiority from human intelligence itself? Does insisting on that distinction help us understand intelligence more precisely — and perhaps produce a better definition of it?
PB: This is a good example of what I mean above. Obviously, computers and AI systems can do many computational tasks far better and faster than we can. This is no surprise, and of course was to some extent true long before the current AI systems. “Intelligence” is a much- and long-contested term, even when we restrict the discussion to humans. It seems clear now that it would be mistaken to imagine the “intelligence” of other animals to be like that of humans but inferior. In animal cognition, intelligence is increasingly recognized to be a multidimensional attribute. Similarly, within the Space of Possible Minds, machine “intelligence” occupies a quite different domain to human, and so it should be obvious that it makes little sense to measure one relative to the other.
I believe that the notion of “artificial general intelligence” (AGI) is incoherent. When pushed, most definitions of AGI are absurdly rigged: basically, it becomes all those aspects of human intelligence that can be measured and quantified, or indeed monetized. Some AI researchers seem almost affronted when one asks why AGI does not include qualities such as empathy, emotional intelligence, or intuition – qualities that are absolutely vital in human cognition. “AGI” is a concept that seems to be defined specifically in a way that makes it a feasible target for AI – something that tech companies can plausibly declare they have achieved a few years from now.
But we should not overlook too the fact that AI as a field has long had a grotesque notion of human intelligence – as we can see from the fact that for a long time the goal that would supposedly demonstrate human-like capabilities was the ability to beat the best human players at chess. When it comes to “AGI”, little has changed.
CH: Current AI systems are often described as processing information without understanding, in the sense that they are not embodied like humans and lack the capacity for social inference. However, humans only infer the existence of understanding indirectly in other humans. If understanding is never directly observable even in human beings, what principled criterion ultimately allows us to deny it in machines?
PB: First, parsimony. It makes sense to suppose that other people have an interior world much like ours, because the alternative is that somehow behaviour and understanding that is just like ours is being produced by an utterly different process.
Some might say: But doesn’t AI today now also show behaviour and understanding that is in some respects indistinguishable from ours? That claim is contentious in the first place, but even if it were true, the argument is specious: we share a common descent and physical makeup to other humans, so it is natural to suppose we share attributes like understanding in common, because we work in the same way. AI systems do not work like us (claims to the contrary are simply uninformed), and so there is no reason to infer that an apparent similarity in behaviour comes from similarity of mechanism or interiority. The default assumption should be just the opposite.
But it goes beyond this. For LLMs are designed explicitly to mimic human understanding. This is what constantly amazes me about chatbots: AI engineers have spent their efforts perfecting machines such that they are capable of delivering outputs that closely resemble human discourse (and of course we must remember that most transformer-enabled algorithm are domain-specific and are not “human impersonators” at all), and then when they succeed, they act surprised and start hinting (or more) that perhaps these machines are conscious. It is indeed remarkable what the machines are capable of – but only in the sense that they do so well what they were designed to do. The fact that they seem so “spookily human” sometimes is not an unexpected and emergent side-effect, but a consequence of their design. As AI ethicist Shannon Vallor has said, they are mirrors of us – and we are sometimes foolish enough to be deceived by the reflection.
“Some might say this is an unfair judgement of AI ‘creativity’. My own view is that, given that (for LLMs anyway) the ‘creativity’ is parasitic on ours, we don’t owe it our appreciation!”
CH: One of the interesting notions presented in "Machine Minds" was that intelligence and consciousness may not necessarily coincide, and that highly capable systems could exist without subjective experience. If that separation is possible, do you think humans systematically overestimate the importance of consciousness when evaluating intelligence?
PB: I’m honestly not sure if we overestimate or underestimate it! The fact is that we still don’t know what consciousness is “for”: why it evolved, and what evolutionary purpose, if any, it serves. There are plenty of ideas about that, some of them quite plausible. For example, some say that conscious awareness injects a kind of piquancy and urgency into decision-making, supplying a motivation for us to attend to what really matters. I’m not wholly convinced that isn’t a rather circular argument though. My hunch is that indeed there was some adaptive benefit to having consciousness, but I haven’t seen a totally compelling case for what this might be. It seems at least plausible that access consciousness – the availability to our awareness of thoughts, understanding, and reasoning – can improve the quality of decision-making. It’s less clear what purpose phenomenal consciousness – why experiences feel a certain way – might have, and indeed some philosophers of mind argue that this is an illusion (not a good word, in my view, but I think I understand the point).
At any rate, I do think it is heuristically useful to think of Mindspace as having directions in “intelligence” in the sense of ability to reason and to problem-solve (as computers can) and in “consciousness/awareness/affect”, which is an independent attribute. Computers have none of the latter (in my view); babies have plenty, but not yet much “intelligence” in the sense of being able to act on the world (except involuntarily, because who can resist a baby?). But I don’t believe consciousness itself is a single thing – it too has several, perhaps many, dimensions. I don’t believe that the consciousness of animals, especially for example octopuses, is merely like ours but lesser.
CH: Across some of your books, you examine systems in which complex behavior emerges through interaction rather than from a single controlling center. Does the language of “mind” risk encouraging us to imagine intelligence as internally centralized, when it may instead emerge relationally through interaction, inference, and environment?
PB: No, I think there are spaces in the Space of Possible Minds for collective intelligence. Collective intelligence can problem-solve, for example, and can be responsive and adaptive. Of course, even the human mind has a strong degree of collectivity about it, as it arises from the interactions of many neurons and of several neural modules, some of which can be in conflict with others. I’d suggest that we think about these collective aspects of mind or intelligence in terms of agency: does the collective act like a single agent, with unity of purpose and behaviour, or do its components have their own agendas and conflicts? It’s not an either/or – there can be an interesting middle ground (and some researchers think that the apparently distributed mind of octopuses might be like this).
“As AI ethicist Shannon Vallor has said, they are mirrors of us – and we are sometimes foolish enough to be deceived by the reflection.”
CH: In the section “Passing the test,” you described how AI-generated poems were rated highly until their source was disclosed, and how Iamus’s compositions were judged comparable to human modernist work until attributed to a machine. Then you wrote, “What’s really striking about AI poetry, prose, visual art, and music is not so much how impressively human-like they are, but how impressed we are by it.” What were people responding to before disclosure intervened? Does attribution itself fundamentally reorganize aesthetic judgement?
PB: Let’s take music, for example. We respond to many aspects of it, generally in terms of patterns and regularities. That’s how we perceive melodies: as organized features in sequences of notes. And we respond to those by generating expectations of what will come next, and our emotions are engaged according to whether our expectations are fulfilled or not.
So we can find such patterns and expectations – we can have a musical experience, you might say – from any non-random sequence of notes. Some slowed-down birdsong can sound amazingly musical to us, because it shares with human music some features like repetitive structures and non-random pitch intervals. It’s absolutely no surprise, then, that AI systems – whether they are LLMs or not – that are trained/engineered to create music-like outputs can elicit human emotion. If they are, say, trained on Chopin, this is even less surprising: they are then just mimicking Chopin’s choices in many musical dimensions (e.g. melodically, harmonically, rhythmically…). This is what we respond to in such computer-generated music: the “music” activates all the mental templates we have developed in listening to music. In this sense, AI “music” ought to remind us that we play a real and even creative role in the musical experience: we’re not just passive recipients.
If/when people seem less enthusiastic about music, poetry, art, when they know it is AI-generated, this is not surprising either. We seek meaning in each other’s creative works, even if we are not sure what meaning is there or if we each have a unique experience. So it stands to reason that we should sense some emptiness when we know there is no such meaning to be found – when we know there is nothing and no one that is striving to convey anything in the artwork. Some might say this is an unfair judgement of AI “creativity”. My own view is that, given that (for LLMs anyway) the “creativity” is parasitic on ours, we don’t owe it our appreciation!
CH: On the judgement of creativity, you suggested that we may need new vocabulary to acknowledge “creative” outputs that are different from ours. Do you think the contemporary debates around AI-generated art are ultimately disputes about the works themselves, or about the adequacy of human-centered categories for recognizing and evaluating cultural production?
PB: This is an excellent but complex question, and I feel like it needs a philosopher of aesthetics to do it justice! I’ll say this much: at this point, I think the debates generally are about the work itself, because it is usually not very good. AI-generated rock and pop music is horrible – not because it is unlistenable (which used to be the case) but because it is so utterly predictable. I have played in several rock bands (amongst other sorts of music), and AI rock makes me blush with shame because it shows what terrible clichés emerge from this averaging of all of rock. It’s literally music-by-numbers, lacking even the slightest spark of originality (and that, of course, by design). I think the same is true for AI-generated classical music, visual art, etc. AI prose fiction is typically dreary drivel, and equally hackneyed. I daresay it will get better, but I do wonder by how much.
But this doesn’t mean that I reject any creative uses of AI. I was intrigued by the way the Iamus AI music system was being used – this was before LLMs, and it used a kind of evolutionary algorithm that didn’t rely on plagiarism. One composer I spoke to was using it as a source of little nuggets of ideas that he would then develop. This seems legitimate to me, and indeed entirely comparable to the way I will sometimes hear a song and say to myself “I can use a bit of that, but in a way that the source will be unrecognizable.” That, I think, is how human creativity has always worked.
“AI ‘music’ ought to remind us that we play a real and even creative role in the musical experience: we’re not just passive recipients.”
CH: "Machine Minds" appeared in The Book of Minds in 2022, before LLMs entered widespread public use and the rapid acceleration of contemporary AI discourse. What, if anything, would you revise or add to the chapter? Do you think the field is still asking the questions the chapter opened with, or have those questions been quietly displaced by others?
PB: I would now need to revise that chapter substantially! But what I’ve said above gives a fair indication of some of the things I’d now need to discuss. I think, for example, that it has become much more challenging to think clearly about the nature of what “machine minds” might be like – we have even more reason to believe that such minds, if they ever truly warrant that term, could work very differently from ours. And there is now much more discussion of what distinctions exist between our minds and the way today’s AI works. One of the things that concerns me most is how we are being encouraged to elide those differences – and not because AI is getting more human-like, but because we are being repeatedly told, sometimes by AI experts, that after all our minds don’t work so differently from LLMs. Well, they really do – and we forget that at our peril.
This conversation is part of Axis of Culture’s Summer 2026 issue exploring how recent developments in artificial intelligence are reshaping our worldviews, and how we understand minds, cognition, and cultural production.Chennie Huang is the Founder of Axis of Culture.