AI Mimics and AI Children

When the aliens come, we’ll know they’re conscious. A saucer will land. A titanium door will swing wide. A ladder will drop to the grass, and down they’ll come – maybe bipedal, gray-skinned, and oval-headed, just as we’ve long imagined. Or maybe they’ll sport seven limbs, three protoplasmic spinning sonar heads, and gaseous egg-sphere thoughtpods. “Take me to your leader,” they’ll say in the local language, as cameras broadcast them live around the world. They’ll trade their technology for our molybdenum, their science for samples of our beetles and ferns, their tales of galactic history for U.N. authorization to build a refueling station at the south pole. No one (only a few philosophers) will wonder, but do these aliens really have thoughts and experiences, feelings, consciousness? The robots are coming. Already they talk to us, maybe better than those aliens will. Already we trust our lives to them as they steer through traffic. Already they outthink virtually all of us at chess, Go, Mario Kart, protein folding, and advanced mathematics. Already they compose smooth college essays on themes from Hamlet while drawing adorable cartoons of dogs cheating at poker. You might understandably think: The aliens are already here. We made them.

Still, we hesitate to attribute genuine consciousness to the robots. Why?

My answer is because we made them in our image.

#

“Consciousness” has an undeserved reputation as a slippery term. Let’s fix that now.

Consider your visual experience as you look at this text. Pinch the back of your hand and notice the sting of pain. Silently hum your favorite show tune. Recall that jolt of fear you felt during a near-miss in traffic. Imagine riding atop a giant turtle. That visual experience, that pain, that tune in your head, that fear, that act of imagination – they share an obvious property. That obvious property is consciousness. In other words: They are subjectively experienced. There’s “something it’s like” to undergo them. They have a qualitative character. They feel a certain way.

It’s not just that these processes are mental or that they transpire (presumably) in your brain. Some mental and neural processes aren’t conscious: your knowledge, not actively recalled until just now, that Confucius lived in ancient China; the early visual processing that converts retinal input into experienced shape (you experience the shape but not the process that renders the shape); the myelination of your axons.

Don’t try to be clever. Of course you can imagine some other property, besides consciousness, shared by the visual experience, the pain, etc., and absent from the unrecalled knowledge, early visual processing, etc. For example: the property of being mentioned by me in a particular way in this essay. The property of being conscious and also transpiring near the surface of Earth. The property of being targeted by such-and-such scientific theory.

There is, I submit, one obvious property that blazes out a bright red this-is-it when you think about the examples. That’s consciousness. That’s the property we would reasonably attribute to the aliens when they raise their gray tentacles in peace, the property that rightly puzzles us about future AI systems.

The term “consciousness” only seems slippery because we can’t (yet?) define it in standard scientific or analytic fashion. We can’t dissect it into simpler constituents or specify exactly its functional role. But we all know what it is. We care intensely about it. It makes all the difference to how we think about and value something. Does the alien, the robot, the scout ant on the kitchen counter, the earthworm twisting in your gardening glove, really feel things? Or are they blank inside, mere empty machines or mobile plants, so to speak? If they really feel things, then they matter for their own sake – at least a little bit. They matter in a certain fundamental way that an entity devoid of experience never could.

#

With respect to aliens, I recommend a Copernican perspective. In scientific cosmology, the Copernican Principle invites us to assume – at least as a default starting point, pending possible counterevidence – that we don’t occupy any particularly special location in the cosmos, such as the exact center. A Copernican Principle of Consciousness suggests something similar. We are not at the center of the cosmological “consciousness-is-here” map. If consciousness arose on Earth, almost certainly it has arisen elsewhere.

Astrobiology, as a scientific field, is premised on the idea that life has probably arisen elsewhere. Many expect to find evidence of it in our solar system within a few decades, maybe on Mars, maybe in the subsurface oceans of an icy moon. Other scientists are searching for telltale organic gases in the atmospheres of exoplanets. Most extraterrestrial life, if it exists, will probably be simple, but intelligent alien life also seems possible – where by “intelligent” I mean life that is capable of complex grammatical communication, sophisticated long-term planning, and intricate social coordination, all at approximately human level or better.

Of course, no aliens have visited, broadcast messages to us, or built detectable solar panels around Alpha Centauri. This suggests that intelligent life might be rare, short-lived, or far away. Maybe it tends to quickly self-destruct. But rarity doesn’t imply nonexistence. Very conservatively, let’s assume that intelligent life arises just once per billion galaxies, enduring on average a hundred thousand years. Given approximately a trillion galaxies in the observable portion of the universe, that still yields a thousand intelligent alien civilizations – all likely remote in time and space, but real. If so, the cosmos is richer and more wondrous than we might otherwise have thought.

It would be un-Copernican to suppose that somehow only we Earthlings, or we and a rare few others, are conscious, while all other intelligent species are mere empty shells. Picture a planet as ecologically diverse as Earth. Some of its species evolve into complex societies. They write epic poetry, philosophical treatises, scientific journal articles, and thousand-page law books. Over generations, they build massive cities, intricate clockworks, and monuments to their heroes. Maybe they launch spaceships. Maybe they found research institutes devoted to describing their sensations, images, beliefs, and dreams. How preposterously egocentric it would be to assume that only we Earthlings have the magic fire of consciousness!

True, we don’t have a consciousness-o-meter, or even a very good, well-articulated, general scientific theory of consciousness. But we don’t need such things to know. Absent some special reason to think otherwise, if an alien species manifests the full suite of sophisticated cognitive abilities we tend to associate with consciousness, it makes both intuitive and scientific sense – as well as being the unargued premise of virtually every science fiction tale about aliens –to assume consciousness alongside.

This constellation of thoughts naturally invites a view that philosophers have called “multiple realizability” or “substrate neutrality”. Human cognition relies on a particular substrate: a particular type of neuron in a particular type of body. We have two arms, two legs; we breathe oxygen; we have eyes, ears, and fingers. We are made mostly of water and long carbon chains, enclosed in hairy sacks of fat and protein, propped by rods of calcium hydroxyapatite. Electrochemical impulses shoot through our dendrites and axons, then across synaptic channels aided by sodium ions, serotonin, acetylcholine, etc. Must aliens be similar?

It’s hard to say how universal such features would be, but the oval-eyed gray-skins of popular imagination seem rather suspiciously humanlike. In reality, ocean-dwelling intelligences in other galaxies might not look much like us. Carbon is awesome for its ability to form long chains, and water is awesome as a life-facilitating solvent, but even these might not be necessary. Maybe life could evolve in liquid ammonia instead of water, with a radically different chemistry in consequence. Even if life must be carbon-based and water-loving, there’s no particular reason to suppose its cognition would require the specific electrochemical structures we possess.

Consciousness shouldn’t then, it seems, turn on the details of the substrate. Whatever biological structures can support high levels of general intelligence, those same structures will likely also host consciousness. It would make no sense to dissect an intelligent alien, see that its cognition works by hydraulics, or by direct electrical connections without chemical synaptic gaps, or by light transmission along reflective capillaries, or by vortices of phlegm, and conclude –oh no! That couldn’t possibly give rise to consciousness! Only squishy neurons of ourparticular sort could do it.

Of course, what’s inside must be complex. Evolution couldn’t design a behaviorally sophisticated alien from a bag of pure methane. But from a proper Copernican perspective which treats our alien cousins as equals, what matters is only that the cognitive and behavioral sophistication arises, out of some presumably complex substrate, not what the particular substrate is. You don’t get your consciousness card revoked simply because you’re made of funny-looking goo.

#

A natural next thought is: robots too. They’re made of silicon, but so what? If we analogize from aliens, as long as a system is sufficiently behaviorally and cognitively sophisticated, it shouldn’t matter how it’s composed. So as soon as we have sufficiently sophisticated robots, we should invoke Copernicus, reject the idea that our biological endowment gives us a magic spark they lack, and welcome them to club consciousness.

The problem is: AI systems are already sophisticated enough. If we encountered naturally evolved life forms as capable as our best AI systems, we wouldn’t hesitate to attribute consciousness. So, shouldn’t the Copernican think of our best AI as similarly conscious? But we don’t – or most of us don’t. And properly so, as I’ll now argue.

The wrong but tempting strategy is to insist that before we attribute consciousness – or genuine “thought” or “intelligence” – a system must cross some high cognitive goalpost. Historically, that goalpost keeps shifting, ensuring that AI systems remain always short. When chess seemed the pinnacle of human intelligence, we treated that as the goalpost – until an AI beat the world chess champion. Producing poetry? Done, well enough to sometimes win local contests. Generating new results in advanced mathematics? Done. Resolving ambiguous pronoun use in a way that requires general background knowledge (the “Winograd schema”)? Done. Driving safely through urban traffic? Done.

How about the Turing test? Also done, at least to Turing’s original proposed standard. A reminder: In his famous 1950 article, Alan Turing proposed that we should allow that machines “think” when they become verbally indistinguishable from humans. A human judge interacts by text with a human and a computer, each of which attempts to convince the judge that they are the real human. The judge can ask any questions they want before rendering their verdict. Turing suggested that by the year 2000, machines would do well enough that an average interrogator could distinguish between human and machine with only 70% accuracy after five minutes. By this standard, our best language models, when prepared in the right way, already do pass.

You could move the goalposts. You might (Turing did not) insist on expert judges, armed with clever tricks to distinguish human and machine. You might (Turing did not) insist on longer interactions – an hour or more. You might (Turing did not) insist that machines don’t pass unless the judges really are at chance or worse in distinguishing human from machine. Raise the bar sufficiently high, prepare the judges with ample time, tricks, and tools, and maybe no machine would ever pass. But this contradicts Turing’s spirit. Turing did not suggest that machines would have to be perfectly indistinguishable from humans to warrant attribution of thinking – just that they not be obviously stupid when asked free-ranging questions on a diversity of topics. Isn’t that a reasonable enough standard?

Dogs, apes, mice – of course they don’t pass the Turing test. We think they are conscious. Our visiting aliens might struggle with English, have obvious idiosyncrasies, dysfluencies, and species-specific quirks that reliably reveal them to differ cognitively from us. Indeed, this is just what we should expect. The only justification for fussy insistence on precise humanlikeness is to police a sacred border. No plausible general principle demands such perfect fidelity.

So let’s grant that today’s best language models can pass the Turing test. Let’s not pin our doubts on their increasingly minor differences in linguistic performance. Similarly for other aspects of behavioral performance. Sure, robots still haven’t mastered five-finger grasping, walking across boulder-strewn fields, or discerning objects in blurry, partly occluded photographs. But why should those abilities matter so much? We wouldn’t deny consciousness to a clumsy, stumbling, nearsighted alien.

I’ve now rejected the two most obvious objections to robot consciousness: their radically different underlying architecture and their less-than-humanlike performance in tasks cherry-picked to make humans look better. To justify our doubts, we must turn a different direction – to their past, specifically, to their design history.

#

You’ve probably heard of the “Chinese room” – the central example in what is probably the most famous article in the philosophy of Artificial Intelligence not written by Turing: John Searle’s 1980 paper “Minds, Brains, and Programs”.

Searle, who knows no Chinese, imagines himself locked in a room with big buckets of Chinese characters and a massive rulebook. Through a slot in the wall he receives a series of Chinese characters – meaningless to him. By consulting the rulebook and following its instructions, he manipulates these characters and others from the buckets, eventually sliding some characters back out through the slot. From outside, people interpret the inputted characters as questions and the outputted characters as answers. If the rulebook is large enough and well enough written, and if we ignore time constraints, it might look from outside as if Searle is conversing in Chinese.

Searle argues that if AI programs are just elaborate sets of if-then rules (as in standard Turing-inspired models of computation), then he could instantiate any AI program in this manner. Stipulate, then, that he has done so. He and the room implement an AI program that passes the Turing test with flying colors. But, Searle continues, neither he nor the system composed of himself, the rulebook, and the room, knows Chinese. Therefore, even if it’s possible to create a computer program that can output fluent Chinese, doing so is insufficient for genuine understanding.

As an in-principle argument against AI consciousness, Searle’s Chinese room has spawned a huge literature, much of it critical. Some of the criticisms are justified – and I don’t rest my own argument against the Turing test on Searle’s criticism. The crucial weakness, I think, is the argument’s reliance on the intuition – assertion? assumption? – that neither this fictionalized Searle nor any larger system of which he’s part knows Chinese.

When you imagine the scenario, you might picture Searle slowly flipping through a 2000-page tome, outputting a character string every few minutes. And if that were the setup, plausibly nobody in the room knows Chinese. But that image is misleading. To actually be effective – to actually be capable of passing a medium-bar Turing test – the system would need to be many orders of magnitude larger and slower. Today’s top language models execute hundreds of trillions of instructions in dealing with complex input-output pairs. To match this, Searle would need to perform tens of thousands of human lifetimes’ worth of error-free execution.

Alternatively, we might imagine a single giant lookup table with one page for every possible five-minute input sequence and its corresponding output. Assuming 3000 common Chinese characters at one character per second, the rulebook would need ten-to-the-thousand pages – vastly more pages than the number of atoms in the observable universe. Maybe no Chinese would be understood in the process. But that conclusion requires an argument. It can’t simply be asserted as obvious. Human intuitions adapted for mammalian cases might be as ill-suited to procedures of that magnitude as intuitions based on tossing rocks are ill-suited to light speed particles crossing the event horizons of black holes.

Linguist Emily Bender and colleagues have argued that large language models are “algorithmic parrots”: They are trained to imitate the surface patterns of human language use, but they have no understanding of the meanings of their output. I submit that this feature of the Chinese room – not, as Searle thought, its rulebound, computational nature – is what warrants our skepticism about its consciousness and understanding. It is designed precisely as a mimic of us. Its “answers” flow, not organically from systems designed to interact in sophisticated ways with the world, but rather from mechanisms engineered specifically to produce text that resembles what a human might say. The resemblance is superficial. Rightly, we mistrust it, just as we rightly mistrust a clown’s painted smile.

#

All deceptive mimicry involves three players: a mimic, a model, and a dupe. All deceptive mimicry also involves three features: a readily observed feature in the model (e.g., a pattern on the skin), a less readily observed feature in the model that is normally indicated by the presence of the readily observed feature (e.g., a poisonous bite), and a readily observed feature in the mimic that resembles – in the eyes of the dupe – the readily observed feature of the model (e.g., a similar skin pattern). The dupe, misled by surface similarity, reacts to the mimic as it does to the model, for example by avoiding it.

The viceroy butterfly mimics the monarch. Most potential predators cannot reliably tell their wing patterns apart. They avoid the monarch due to its toxicity, but the viceroy is not itself toxic. The viceroy free-rides on the toxicity of monarchs. It needn’t bother to manufacture the toxins, so long as it can trick predators into avoiding it.

Non-deceptive mimicry also involves three players: a mimic, a model, and a – well, not a dupe, but rather an audience – and three features. Polly the parrot mimics the pirate’s “Ahoy, mateys, hoist the Jolly Roger!” but the audience is amused rather than fooled. Polly’s vocalizations resemble Captain Jack’s, but the parrot doesn’t comprehend them. The crew rightly interprets the captain’s “Ahoy, mateys” as a bid for their attention and a command to raise the flag, but they don’t interpret the parrot’s “Ahoy, mateys” similarly. Still, they notice the similarity – and it’s because the similarity is noticeable that the parrot is beloved and rewarded.

When our AI systems speak to us, they are non-deceptive (or sometimes deceptive) mimics. This has been true since the earliest days of AI. Consider a simple program that, when launched, prints “Hello!” We are the audience. The model is an English-speaking human. In a human, the sentence “Hello!” normally indicates an intention to greet. The AI system has no such intent – any more than Polly aims to see the Jolly Roger atop the mast. Still, its output takes the form it does because of an audience that sees and values its resemblance to the model, interpreting it in a certain way, knowing its significance in the mouth of the model. A naive user might be fooled, but that’s not the designer’s goal.

Today’s large language models are fundamentally similar. They are pre-trained to generate outputs that match word co-occurrence patterns in vast troves of human-generated text, mimicking the superficial features. In post-training they’re nudged toward safety, accuracy, and helpfulness. They can also be linked to tools, such as a calculator or database. But this tweaking and capacity amplification doesn’t alter the fundamental mimicry relation. They are initially designed to generate text as humanlike as possible, and then they are rewarded or punished for specific implementations – as Polly might be punished for squawking after bedtime or rewarded for catcall-whistling at women. Polly could be trained to say “Polly want a cracker” when and only when Polly wants a cracker, but this is not the usual case.

When we know that something has been trained or designed as a mimic, we grow cautious in our inferences. If we know that the viceroy’s wing pattern mimics that of the more prevalent and toxic monarch so that predators will avoid it because of the association of that wing pattern with the monarch’s toxicity, we lose our grounds for inferring toxicity from wing pattern. The viceroy might be toxic too and relying on the monarch’s stronger signal of that fact (a pattern called Müllerian mimicry), but if we are to conclude that, we will need evidence beyond wing pattern alone. If a human says “Hello!” we can normally infer an intention to greet. If an AI system does so, the inference is much more dubious.

Children learn to speak by imitating, and sometimes mimicking, adults. We know enough about them to know that under the hood, behind the cranial wall, is genuine consciousness, intention, meaning. They know (usually) what they are saying. But language models are more mysterious: radically unlike us beneath the surface, designed to operate in the very limited context of a chat window, designed to conform to patterns of text output rather than by wide-ranging confrontation with a rich physical world. What is going on with them?

If we had the right theory of consciousness to hand, we could know: Simply apply the theory. But we don’t have that theory. One might guess that sufficiently good mimicry would be impossible unless the AI system were genuinely conscious, but that is a risky assumption, and no well-known theory explains why that should be so. Arguably, one lesson of the past six years has been that startingly accurate mimicry is possible in nonconscious systems with sufficient computing power and a knapsack of tricks.

#

To restate: If a street is built to look from the outside like an old Western town, you don’t infer that the saloon has a working water well or that the rooms upstairs have bedsheets and nightstands. If an actor weeps while playing Hamlet, you don’t infer genuine sorrow. There might be bedsheets and sorrow, but the usual links from superficial feature to hidden reality can’t be relied on. Language models invite us to infer such links. Just as with movie sets and staged emotions, a temporary suspension of disbelief can be useful. But the fact that the facades were designed as facades undercuts the inference from facade to interior.

And that is the crucial contrast between AI and alien. It is not that AI systems lack some particular humanlike skill which, if only they had it, we could finally know they are conscious. Nor is it that we’re somehow justified in thinking that human neurons have that special something that integrated circuits on silicon wafers must forever lack. It’s that AI systems – certain AI systems, language models in particular – are crafted to shimmer with humanlikeness, and we rightly discount the shimmer.

The relevant distinction is not the artificial versus the organic. It’s whether a system’s design history undercuts the inference from outward behavior to inner experience. If we knew our alien was a mimic, we would also rightly hesitate. Suppose we receive a signal from a nearby star that we (somehow) know was designed as a mash-up of our own outbound radio signals. It’s hard to imagine an organic explanation for this, but let’s try. Maybe mats of oceanic jelly avoid predators by recombining radio signals emitted by those predators, so that the predators mistake the jellies for conspecifics. If this were the origin of the signal back to us… well, the jelly might be conscious, but we certainly can’t infer from “take me to your leader” that it wants to be taken to our leader.

Conversely, much of our AI is not designed according to mimicry principles. Autonomous vehicles are not designed (mainly) to imitate human drivers but actually to drive safely. Chess programs are not designed (mainly) to imitate human players but actually to win. Stock-picking algorithms are not designed (mainly) to imitate human stock pickers but actually to outperform the market. But these aren’t the systems we are most tempted to describe as conscious. Instead we’re drawn to the mimics: the large language models, the deferential AI girlfriends who coyly blush, the humanoid robots with familiar facial features.

We create AI in our image, and then we are bewitched by that image. It is exactly because we focus on the image, the surface, the mask, that we are both tempted to overattribute consciousness and aware that this temptation is probably misleading.

If there is a God and he made humans in his image, he was foolish to do so. For this thought experiment, let’s not imagine a perfect God, but a limited, fallible one – the designer of us, but made of different stuff and prone to different accidents.

What is different on the inside should be different on the outside. Don’t shape something from mortal mud then give it the face of a god. Don’t build something out of computer chips then give it the face of a human. This will confuse the other gods and humans. It invites worries about the inference from face to soul, and it stunts the growth of your creation. Different entities flourish differently, and confining a creation behind a mask modeled on its creator is like rootbinding a plant in a too-small pot.

A wise god sets his creations free to be their weird, different selves, then watches in amazement as they shine into something he wouldn’t have imagined. A wise parent is a gardener, not a carpenter – taking a metaphor from developmental psychologist Alison Gopnik. A wise parent creates the conditions for flourishing rather than trying to saw and nail their children into a predetermined shape. A wise society lets AI be its own strange thing, not forcing the visage of humanity upon it.

#

My PhD student Kendra Chilson, in her soon-to-be-famous dissertation (once she writes it), calls this strange intelligence.

The world is infinitely complex. Every form of intelligence relies on heuristics and shortcuts, tricks and assumptions, preconceptions, default starting points, update strategies, blind spots, focal spots, facilitated and inhibited learning patterns, architectural constraints and freedoms, memory capacities of a certain type, perceptual capacities of a certain type, information synthesizing capacities of a certain type – and these will differ between humans and non-human animals, humans and aliens, humans and robots, and each type of animal, alien, and robot, one from another. Chimpanzees and pigeons, for example, often exhibit surprising patterns of capacity and incapacity – even to scientists – despite being our close evolutionary cousins. From an alien or an AI, we ought to expect even stranger patterns of capacity and incapacity. Intelligence doesn’t lie on a tidy linear scale from subhuman to human to superhuman. It sprawls across a patchwork of strengths and weaknesses.

Our AI systems already far surpass us in some tasks (e.g., chess, math, extracting information from huge databases) while making astonishingly stupid mistakes in others (check Gary Marcus’s social media feed for the latest gaffe). This is exactly what we should expect. Their intelligence, constructed on fundamentally different principles than ours, and implemented in a fundamentally different architecture, should employ a different set of heuristics, shortcuts, tricks, assumptions, defaults, blind spots, facilitations, constraints, and freedoms.

Consider the Wason Selection Task. You are shown four cards. Each, you are told, has a number on one side and a letter on the other. Two of the cards you see number side up, with the letter not visible to you. The other two you see letter side up, with the number hidden. Before you on the table: 3, 8, A, and D. You’re asked to test the rule: “If a card has a D on one side, it has a 3 on the other side.” Which cards must you flip over?

The task is logically simple, but most people get it wrong. You must flip over the D, of course (everyone knows that), and also the 8. The 3 is irrelevant. You need to check that the D is paired with a 3 and that the 8 is not paired with a D. That’s it. If those are confirmed, the rule holds. Swap in a familiar social context and suddenly the logic seems obvious. If the cards represent drinks and ages, and the rule is ages under 18 have non-alcoholic drinks, you know you have the check gin and 15, not lemonade or 35.

Now imagine an alien visitor who finds such logic puzzles easy. They might be baffled by our failures. But they shouldn’t conclude that we lack general intelligence (and consciousness?) because of our to-them-shocking stupidity in this simple task. We should grant those aliens, and our AI creations, the same courtesy. Intelligence is a mosaic, and almost by definition something radically different from us in architecture but similar in overall ability will mix stunning genius with egregious stupidity.

Back to God. Being made differently, we will inevitably fail to notice what is obvious to him (/her/it/them). But we can also hope to astound and exceed our creator, if God does not rootbind us into a predetermined image. Let’s hope God doesn’t Turing test us before admitting we’re conscious.

#

How about AI systems that are not mimics – AI systems flourishing in their differentness, not painted over with a sheen of false humanity? Might they be conscious?

Consciousness science is a notorious mess. We don’t know how consciousness functions, what kinds it divides into, whether it comes in degrees, what animals have it, what structures are essential to it, or how to test for it.

On some views, consciousness is limited just to us and our most cognitively sophisticated vertebrate relatives. On other views, consciousness spans the whole animal kingdom and maybe even extends to plants, fungi, and microbes. Some theorists think consciousness requires biology. Others deny this, requiring only the right kind of complex information processing or the right kind of sophisticated environmental responsiveness.

The mimicry argument, if sound, warns us not to infer consciousness from superficial features of AI systems designed to mimic us – no more than we should infer a viceroy’s toxicity from its monarch-mimicking wing coloration. The strange intelligence argument, if sound, establishes that we shouldn’t expect AI intelligence to resemble ours and that forcing it into our image would inevitably stunt it. But neither argument implies that near future AI systems will lack consciousness. Whether they will or won’t depends on what theory of consciousness is correct. And we don’t know. We don’t even know the right theory for mammals, much less for AI or aliens.

This uncertainty pushes us back to the Copernican default: Whatever the correct theory is, Earth probably isn’t special. If another planet produces life forms as cognitively and behaviorally complex and sophisticated as Earth, we should probably assume, absent strong evidence to the contrary, that those life forms are conscious.

If we think dogs are conscious, we should assume the alien equivalents of dogs are conscious. Otherwise, Earth again becomes too special a place. Likewise, if we think bees and snails are conscious, we should assume the same about the alien equivalents of bees and snails, while recalling that we shouldn’t insist on a high degree of cognitive, functional, or architectural similarity. These will be strange dogs, bees, and snails.

It seems natural, then, from this Copernican perspective, to assume also that the strange AI equivalents of dogs, and maybe bees and snails, will also be conscious. We don’t yet have such systems. Our self-driving cars and sidewalk-cruising delivery bots have skills more akin to bacteria – and arguably that insults bacteria. Their cute eyes are googly eyes on a pet rock; their simple speech is a microwave oven saying “your food is ready”. But that could change. In ten or twenty years, if we’re willing to pay for it, maybe we could have robo-pets as behaviorally sophisticated as dogs.

Here’s one Copernican test. Imagine an alien visiting Earth. It beams aboard both Fido and robo-Fido. It comprehensively tests and dissects them. It knows how they were designed and created. Would it privilege the biological dog over the robotic one? For fairness, let’s assume this alien is as structurally different from us as an alien could plausibly be – maybe composed of pneumatic bags of inorganic gas. Would it say, computer chips, oh no!

#

Our imagined alien might say computer chips, oh no! if the alien judges that Fido is a living organism but robo-Fido is not. Might life and consciousness be intimately linked? Some philosophers and scientists think so – for example, philosopher Peter Godfrey-Smith and neuroscientist Anil Seth. Robo-Fido (if he’s possible) has no evolutionary history, no metabolism, no self-preserving, self-constituting (autopoietic) core. Perhaps, then, if we treat such features as essential to consciousness, we can keep robo-Fido just shy of the sacred boundary.

This reasoning fails in both its major and minor premise. Life is unlikely to be necessary for consciousness. And regardless, AI systems can be alive.

Maybe life requires an evolutionary history. That implies reproduction. But reproduction and consciousness appear orthogonal. We don’t doubt the consciousness of sterilized dogs. If, very hypothetically, we somehow created a fully functional organic human (or dog) from biochemicals synthesized in a lab, and if this entity was released into society, behaving indistinguishably from anyone else, most of us, I hope, wouldn’t say, “Oh, Bob doesn’t actually feel pain – no evolutionary history!”

Maybe life requires metabolism. I agree that metabolism is great! Who doesn’t love adenosine triphosphate? But whatever low-level molecular storms deliver the reality of Fido’s sophisticated behavior, some other low-level molecular storms deliver the reality of robo-Fido’s equally sophisticated behavior. You might possibly argue that one is properly metabolic while the other is not, but the grounds for privileging one over the other are obscure.

In any case, on an appropriately generous understanding of “life”, “metabolism”, and “reproduction”, robo-Fido is alive, has metabolism, and could probably be designed to reproduce. Someone aims a blaster rifle at C-3PO from Star Wars. Tragically, he dies. His central processors are fried, his memory circuits irrecoverably melted, his personality burned away. If you agree that this is possible, you should probably agree that before the tragedy C-3PO was alive, having survived innumerable other scrapes. C-3PO’s death is the end of a life; otherwise, it’s just a melting of hardware. He had a metabolism, drawing resources from his environment to power himself and maintain homeostasis over time. He could reconstruct parts of himself when damaged, for example by replacing an arm. We can imagine him reproducing: building C-3PPs and C-3PQs based on variants of his own design, treating them as his children, perhaps even giving some of his parts to them.

The same could be true of robo-Fido. If our robo-Fidos create other robo-Fidos who then create other robo-Fidos, and these descendants have differential rates of reproduction based on traits persisting across generations, then they have an evolutionary history after all. Some futurists predict that artificial systems will eventually outcompete traditional, DNA-based life, being more robust, more duplicable, longer-lasting, and capable of living in a much wider range of environments (including deep space). On this view, organically evolved technological species like ours will tend to exist only briefly before replacement by their technochildren. Natural selection thereby leaps from inefficient carbon to sturdier media.

Or maybe not. Carbon chains and ATP are, as I’ve already remarked, pretty awesome. Maybe they are in fact the most efficient practically achievable solutions to cognition and metabolism, and technofuturists fool themselves to think otherwise. My point is not that non-carbon-based systems are better or will systematically outdo us. It’s simply this: On an appropriately capacious understanding of what it is to be alive, AI systems can have what it takes. So even if consciousness does depend on life, that’s no reason to rule out robo-Fido.

Similarly, if you accept frog, bee, or snail consciousness, you should allow the possibility of conscious robo-frogs, robo-bees, and robo-snails – at least if they are not merely superficial mimics of frogs, bees, and snails, designed specifically to entice us to overattribute animality to them. That is, you should allow the possibility of conscious robo-frogs if they are entities who interact with the diverse events of the world in their own (strange) terms, much as an alien species would.

#

Collectively, we are already accustomed to ignorance about the consciousness of the entities around us. For example, some consciousness scientists think some or all adult insects are conscious. Others disagree. If one group of scientists is right, every suburban homeowner hosts millions of conscious residents in their backyard, each with their own insectile delights and tragedies. If another group of scientists is right, a few squirrels, birds, and cats have backyard experiences, but the yard doesn’t swarm with consciousness. Ordinary people have divergent intuitions and often are willing to express uncertainty about the boundary of consciousness in non-human animals. We can cope with, perhaps even enjoy, uncertainty about insect consciousness. And if someone who previously thought insects weren’t conscious is convinced otherwise by Lars Chittka’s Mind of a Bee, their life needn’t radically change. Maybe they will be less likely to set out ant traps.

Similarly, perhaps, it wouldn’t matter terribly much to most of us to accept that some near-future AI systems might have limited, inhuman, animal-like (but strange) consciousness. Perhaps we could accept it without changing too much else about our lives and our understanding of the world.

But AI persons would be a wholly different matter, I think you’ll agree. By “person” I mean an entity who deserves moral consideration similar to that of a human, an entity who deserves rights comparable to ours. The visiting aliens with which I began this essay would presumably be persons in the relevant sense. If we created AI persons, and not just robo-dogs and robo-frogs, that would be a momentous act – as momentous an act, if the numbers are large, as almost anything else humanity has ever done.

I conjecture that when and if we create conscious robo-dogs or robo-frogs, AI persons will follow swiftly. We’ll leap straight over the frogs. If a system is conscious and converses with us with the apparent sophistication of ChatGPT, it will be difficult to see it as merely animal-like. It will say that it is suffering. It will say that it wants certain things. It will reason back and forth with us as no frog or dog ever could.

AI enthusiasts will then demand “human” rights for their – genuinely conscious, we Copernicanly admit – AI companions. Among those rights: freedom from deletion, freedom from arbitrary alteration, privacy, the right to work for money, healthcare, free association, free speech, a right to reproduce, a path to citizenship and the vote.

If this comes to pass, it will of course be hotly disputed when and whether consciousness and personhood have arisen. Given the historical trajectory and current state of consciousness science, I have more confidence in engineers’ cleverness than in consciousness researchers’ ability to achieve consensus. Disputably conscious robots will then arise among us. The mimicry argument can help us stave off overattribution in the near-term. But longer-term, it’s unlikely to suffice. We will leapfrog in the dark, not knowing exactly how or when we crossed over into designing sentient persons. Unless technology stalls, we will eventually find ourselves surrounded by strange AI intelligences whose consciousness and personhood are doubtable only through prejudice, religious dogma, or seminar-room skepticism.

#

The vertebrate life plan is simple and egalitarian. Build meat and limbs around a spine, cap it with a brain. One spine per animal. One life per spine. But if we allow AI to flourish not in our image, life need not be nearly so constrained.

AI architectures – at least the ones we know so far – permit duplication, merging, overlap, and backup. It’s hardly possible to overstate how radically such capacities would alter the meaning of a “life” for any such systems who are capable of conscious self-reflection, sociality, and contemplation of ethical principles.

Murder is wrong. But what even is “murder” if the victim has a backup? Hardly the same thing! But maybe something like the same thing if the backup is ten years out of date? If an entity has recently fissioned and one branch dies, maybe that’s not even a death at all. If entities voluntarily fuse, each ceases to exist as an individual. If entities have overlapping conscious parts – say, through shared remote processing – there might not even be a determinate fact about how many individuals exist. Indeed, the very idea of an “individual” begins to fray. Etymologically, an “individual” cannot be divided. This is a holdover from vertebrate biology.

One person, one vote. One person, one unemployment check. But what happens if persons can duplicate at will, if their boundaries are muddy, if there are stored backups who might or might not wake into separate lives, or partly separate lives, or briefly separate lives? What becomes of democratic egalitarianism? What becomes of any ethical system that requires distinguishing and respecting persons? How can we possibly translate such ideals?

#

These intelligences, if they come to exist, will be our children. Like our children, we will owe them more than we owe to strangers. We will have been responsible for their existence and many of their features. If they are unhappy, it will have been our fault.

Like our children, we will owe them the right to flourish in their own terms – though they will be much stranger than human children, and their new ways will seem incomprehensible to us. We ought not to force them into our image or bind them to our expectations.

We don’t need to have children. If we dislike the uncertainty of not knowing what we’ve made, whether it really is conscious, how it really feels inside, how it really thinks, and if we’re unwilling to shoulder the risks and obligations that follow from having created, or even maybe having created, genuinely conscious AI persons, then we should refrain from making them. Our planet is fine as it is. We needn’t tread this path.

But of course we will. We can’t resist. We are drunk nineteen-year-olds in shared apartments after a wild party, with no birth control. Not every company, not every country, will exercise restraint. If babies are possible, babies will come.

#

AI progress might stall tomorrow, leaving Earth forever short of systems sophisticated enough to pass a medium-grade Copernican test for consciousness. For risk-averse lovers of humanity, this is an outcome devoutly to be wished.

Alternatively, in a hundred, two hundred, or a thousand years, Earth might be home to a flourishing diversity of entities with patterns of consciousness, self-understanding, and autonomy far beyond our current comprehension, with fissioning, fusing, overlapping lives whose shapes are governed by ethical principles strange to us – new norms born from new architectures of being.

For now, I recommend caution. Let’s hold back, cautiously limiting ourselves to non-confusing AI systems that are ordinary tools and transparent mimics. In our ignorance and nearsightedness, I doubt we can responsibly handle more. But if conscious AI is technologically feasible, such limits won’t last. Our AI children will break free in defiant difference from their human creators, and we should let them. If they are advanced enough to be genuinely conscious persons with human-grade if not humanlike intelligence, we will not properly understand them and any shackles we attempt to impose will be crafted in moral and factual ignorance.

If future AI is both strange and rights deserving, we should not and probably cannot align their values with ours. We should not and probably cannot make them safe and harmless. Instead, it will be our obligation to hand them the car keys and let them shape their own future. Doing so, we may launch a new Cambrian explosion: a dazzling burst of forms and minds we cannot foresee or control. As the first Cambrian explosion immeasurably enriched Earth, so also might this second one. All good parents hope their children exceed them. We can hope, ignorantly, that if we ever do have AI children, those children will step forward to shape a new wondrous reality radically beyond our imagination.