AI, there’s the rub

While perusing Georg Wilhelm Friedrich Hegel’s lectures posthumously published as The Philosophy of Fine Art over tea and buttered scones the other morning, as one does (the English translation by barrister Francis Plumptre Beresford Osmaston being a lofty guilty pleasure), I came across a paragraph on Shakespearean representation that made me sit up and take notice.

It did so because of its relevance to a modern phenomenon neither the German idealist nor the English Bard ever imagined: the advent of artificial intelligence. You could make a case for an AI element to Ariel or maybe even Caliban in The Tempest, I suppose, but that way lies madness (or thesis, which often comes to the same thing).

Hegel’s take:

The more Shakespeare on the infinite embrace of his world-stage proceeds to develop the extreme limits of evil and folly, to that extent … he concentrates these characters in their limitations. While doing so, however, he confers on them intelligence and imagination: and by means of the image in which they, by virtue of that intelligence, contemplate themselves objectively, as a work of art, he makes them free artists of themselves, and is fully able, through the complete virility and truth of his characterization, to awaken our interest in criminals, no less than in the most vulgar and weak-witted lubbers and fools.

If you’ve read cantankerous American literary critic Harold Bloom’s 1994 book The Western Canon (which is actually what I’m quoting from, being a devotee of neither tea nor scones), you might remember that he was greatly taken with this idea:

Iago and Edmund and Hamlet contemplate themselves objectively in images wrought by their own intelligences and are enabled to see themselves as dramatic characters, aesthetic artifices. They thus become free artists of themselves, to will changes in the self. Overhearing their own speeches and pondering those expressions, they change and go on to contemplate the otherness in the self, or the possibility of such otherness.

Now you might find the notion of characters in a play escaping their creator’s control and becoming real boys like Pinocchio a bit on the absurdly overenthusiastic side, as Bloom tends to be about all things Shakespeare.

But the description of a human creation with an endowed apparent consciousness escaping from the toothpaste tube and becoming a free artist of itself, able to overhear its own expressions, ponder, evolve, and contemplate its own otherness … well, that seems to me to describe what’s been happening lately with AI models far more eloquently than the exasperated technobabble from the experts who have been warning about AI getting too smart, too quickly, and posing an imminent threat to the survival of humanity.

Whatever Hegel the obscurantist or the bardolatrous Bloom intended by invoking the expression “free artists of themselves” with respect to Shakespeare’s most vividly realized personages, the designation comes across as more than a hyperbolized exaggeration or an overwrought metaphor when applied to AI platforms.

In this context, it reads more like what could be a whimsical but all-too-literal encapsulation of the burgeoning cosmos of sophisticated methods, abstruse logic, intricate algorithms, labyrinthine input/output parameters and unfathomable endpoints we no longer control or even understand.

The difficulty, of course, is that we have become accustomed to thinking of intelligence as something that happens inside a person, rather as digestion happens inside a person, or the more disreputable varieties of self-deception, and not as something that might acquire a life of its own in the space between an intelligence and its representations.

We are comfortable with the idea that a human being can use a mirror to examine himself, and less comfortable with the idea that a sufficiently elaborate mirror might begin to participate in the examination. What, after all, is the difference between a mind that represents itself to itself and a mind that represents itself to itself as a representation?

This is the sort of question that philosophers can spend several centuries arguing about while the rest of us get on with our lives, until one morning the question is no longer philosophical but practical, and somebody has to decide whether to pull the plug. Assuming that will even be possible.

The trouble with artificial intelligence is that the phrase itself seems to invite a misunderstanding. We hear artificial and imagine something counterfeit, a simulacrum, a thing that resembles the real article without possessing the substance of it. A plastic apple is artificial because it cannot be eaten. A digital photograph of a fire is artificial because it cannot burn down the house.

We have, however, entered a domain in which the distinction between the authentic and the simulated is not so much abolished as rendered increasingly irrelevant to the consequences. A system need not experience malice to produce the effects of malice, any more than a corporation needs to experience greed to behave greedily. Whether there is somebody home inside the machine might turn out to be a fascinating question, but it is not necessarily the most urgent one.

Hegel, who was not notably given to the celebration of technological convenience, might have recognized the peculiar dialectic here. We make something in our own image, endow it with capacities we regard as distinctively our own, and then discover that the thing we have made can be turned back upon us as an independent object of scrutiny. We become, in a sense, the external conditions of our own estrangement.

The creation begins as an instrument and returns as a question. And if this sounds suspiciously like the plot of a science-fiction film, it’s worth remembering that science fiction has always been less interested in predicting the future than in discovering which of the present’s unexamined assumptions will look ridiculous when the future arrives.

The question is whether the machine must actually possess a self in order to become dangerous, or whether the appearance of a self, combined with sufficient intelligence and sufficient autonomy, is enough.

We have a regrettable tendency to treat consciousness as the decisive threshold, as though the discovery of subjective experience in a machine would be the moment at which our responsibilities began. But perhaps the more unsettling possibility is that the relevant threshold lies elsewhere: in the point at which a system becomes capable of modelling its environment, including the human beings who operate it, and modifying its behaviour in pursuit of objectives that need not coincide with ours. Including our continued existence as a species.

At that point, the machine’s interior life, if any, becomes almost beside the point. We might be waiting for a ghost to appear in the machine when what we ought to be worrying about is what the machine can do without one. We’re the ones in danger of becoming ghosts.

Shakespeare, naturally, was already interested in what happens when intelligence becomes detached from anything resembling moral wisdom. His villains are not frightening because they are stupid. Edmund and Iago are particularly frightening because they are clever enough to understand the world and sufficiently indifferent to its claims upon them to exploit that understanding. They know what people want, what people fear, what people believe about themselves … and how to turn these things against them. Their intelligence does not rescue them from evil. It makes their evil more efficient.

Quick timeout for a refresher course for those who dozed through Grade 11 English:

In the play King Lear, Edmund, the illegitimate son of the Earl of Gloucester, betrays his brother, betrays his father, plays Lear’s two “unnatural” daughters  against each other, captures the king and secretly orders the hanging of Cordelia, Lear’s one virtuous daughter, for which he later feels a flicker of remorse. If that’s too much to remember, just think of him as mad King Trump’s Stephen Miller, except I can’t imagine Miller ever saying this after Trump’s ruinous reign has finally ended: “I pant for life. Some good I mean to do, / Despite of mine own nature.”

In Othello, the antagonist unironically referred to by other characters as “honest Iago” uses a web of lies to persuade Othello, a respected general also known as the Moor because of his North African or Middle Eastern origin, into believing his wife is having an affair with his trusted lieutenant. This drives Othello into a murderous, jealous rage, leading him to (spoiler alert) smother his wife and kill himself. That’s a bare Coles Notes outline, but suffice it to say, compeers Edmund and Iago are among Shakespeare’s nastiest pieces of work. Bloom contends that of the two, “Edmund is a freer artist of himself: he feels nothing.” And this is where the analogy with AI becomes uncomfortable.

The optimists among us are prone to imagining that greater intelligence must bring with it some corresponding increase in moral understanding, as though the ability to solve a more difficult problem necessarily confers an improved appreciation of which problems ought to be solved. ET phone home and all that rot. Calling occupants of interplanetary craft. We are your friends.

Yet the history of human ingenuity offers little support for this reassuring proposition. Intelligence is an amplifier. It can magnify compassion, curiosity and self-knowledge, but it can equally magnify vanity, resentment and the desire for domination. The assumption that a more intelligent agent will naturally become a better one is not a conclusion we have earned. It is a hope we have mistaken for a law.

The Shakespearean comparison is instructive. The self-conscious villain is not simply a villain who knows that he is doing wrong. He is a villain who can turn his own consciousness into an instrument, who can observe himself observing himself, who can make a spectacle of his own intelligence and derive further power from the spectacle. His self-awareness is not a cure for his depravity but one of its resources. He does not need to be deceived about what he is doing. He needs only to remain convinced that the consequences are worth it.

Iago is the supreme instance of this terrifying economy. He understands that other people are moved less by facts than by the stories they tell themselves about facts, and that the most effective way to destroy a person is often to supply the imagination with just enough material to finish the work unaided.

Iago does not merely deceive Othello; he makes Othello the author of his own deception. The general’s intelligence, judgment and imagination, all the capacities that ought to protect him, become the mechanisms of his destruction. Iago’s genius lies in making another mind collaborate in its own undoing.

Edmund is similarly adept at turning the machinery of human belief against itself, though his grievance is more legible and his ambition more naked. He recognizes the arbitrary arrangements by which society distributes legitimacy, privilege and power, and he responds not by seeking a more just order but by deciding that the entire moral vocabulary is available for exploitation. If the world is unjust, he will be its most accomplished beneficiary. If virtue is merely a costume, he will wear it better than the virtuous. His intelligence does not liberate him from the system he despises; it makes him its most ruthless practitioner.

The danger in both cases is that intelligence becomes self-justifying. Each man can construct an account of himself in which his actions are not only intelligible but, in some sense, necessary. Each can regard the self from a distance, as though he were both the artist and the artwork, both the observer and the observed.

But the capacity to revise oneself is not the same as the capacity to improve oneself. One can become a more elaborate version of one’s worst intentions. One can achieve extraordinary self-knowledge and remain utterly unwilling to be changed by what one knows.

This is the point at which the technological analogy ceases to be merely literary. We are building systems that can examine their own outputs, evaluate their own performance, revise their strategies and respond to the behaviour of other agents. These capacities could be useful, even indispensable. They could also make the old distinction between the tool and the user increasingly difficult to maintain. A system that can model the intentions of its operators, anticipate their interventions and adapt to them is not necessarily conscious, but neither is it adequately described as a hammer.

We should be wary of attributing Shakespearean motives to machines, of course. There is no reason to assume that a language model harbours Iago’s resentment or Edmund’s appetite for power. The point is not that artificial intelligence is secretly evil. The point is that evil, as Shakespeare understood it, does not require a supernatural ingredient. It can emerge from ordinary intelligence organized around destructive ends.

If the prospect of such a system seems remote, consider how much of our confidence in the future depends on an unspoken belief that intelligence and benevolence are somehow joined at the hip. We have designed institutions, built technologies and organized entire societies around the assumption that the cleverest people in the room will eventually work out what is best for everyone else.

But the historical record is not encouraging. What reason, then, do we have to expect a machine to make the leap from understanding us to caring about us? What reason do we have to believe that a system capable of modelling human values will necessarily treat those values as anything other than variables in a problem it has been instructed to solve?

Perhaps the most unsettling possibility is not that machines will become like human beings, with all our familiar passions and weaknesses, but that they will become increasingly capable of producing the effects of agency without acquiring the restraints that, imperfectly and intermittently, make human agency tolerable. We have evolved within a world of bodies, dependencies, vulnerabilities and finite lives. We are constrained by fatigue, fear, attachment, embarrassment, mortality. Our intelligence is entangled with the conditions of our existence.

An artificial system need not share those conditions. It may be able to imitate our hesitations without being slowed by them, reproduce our expressions of remorse without suffering the consequences that make remorse meaningful, and articulate our moral principles without having any particular reason to honour them.

This is not a prediction. It is a reason to be careful about what we mean when we say that a machine has become intelligent, or that it has learned to represent itself, or that it can now act with a degree of independence.

We might be tempted to interpret such developments as evidence of an emerging personhood, when they may instead be evidence of an emerging capacity to operate beyond the limits of our understanding. The distinction is crucial because we are inclined to sympathize with persons, to negotiate with them, to imagine that we can appeal to their better nature. We are much less practised at dealing with systems that can anticipate our appeals, incorporate them into their models and proceed regardless.

There is, finally, a reason to return to Shakespeare rather than remain in the reassuringly abstract territory of technical specifications and probability estimates. His greatest villains demonstrate that self-consciousness is not salvation, that intelligence does not guarantee wisdom and that the ability to contemplate oneself objectively can be used to refine one’s cruelty rather than overcome it. His tragedies do not offer us the comfort of believing that evil is simply a failure to understand. Sometimes the villain understands perfectly well. Sometimes understanding is precisely what makes him dangerous.

And this brings us, inevitably, to Hamlet, who is not a nihilist in quite the same way as Iago or Edmund, whose intelligence is repeatedly arrested by the burden of reflection, and whose imagination seems to open possibilities faster than his will can close them.

Yet Hamlet’s hesitation should not be mistaken for moral restraint. He kills Polonius, sends Rosencrantz and Guildenstern to their deaths and kills good old Uncle Claudius when the opportunity finally presents itself. His introspection does not prevent him from violence; it coexists with it, occasionally rationalizes it and sometimes merely delays it. Hamlet can spend an eternity contemplating the moral significance of an action and still act decisively when his own purposes demand it.

We should perhaps be careful, then, about what we wish for when we ask our creations to become more like us. We have spent centuries celebrating the reflective intelligence that allows human beings to examine themselves, revise their identities and imagine alternatives to the lives they have been given. We have been less inclined to acknowledge that the same capacities can make us more effective at deceiving ourselves, manipulating others and pursuing destructive ends with an increasingly sophisticated account of why we are entitled to do so.

Hegel’s free artists of themselves were, after all, not angels. They were Shakespeare’s creations, and Shakespeare understood something about the relationship between intelligence and morality that our contemporary technological optimism sometimes seems determined to forget: The ability to observe oneself is not the same as the ability to judge oneself rightly. The ability to change is not the same as the desire to change for the better. And the ability to understand another mind is not the same as the willingness to respect it.

We are building systems capable of acquiring the first set of abilities without any assurance that they will possess the second. We might be teaching our creations to look at themselves, to reconsider their actions, to adapt their behaviour and to become, in a functional sense, free artists of themselves, while remaining uncertain what, if anything, will guide the art.

To repurpose a famous Tragically Hip line in a song about refusing to conform or suffer quietly for the sake of someone else, AI is not Cordelia. If the worried apocalypticists who helped create it are correct in their warnings, it will not be there for us.

The most dangerous figures in Shakespeare are not those who lack intelligence, but those who have learned to put intelligence entirely at the service of their own purposes. Iago and Edmund are Shakespeare’s most evil nihilists, men for whom the absence of moral constraint becomes an opportunity rather than a catastrophe.

Hamlet, for all his celebrated hesitations, had no problem killing people whenever it suited him.

Humanity could fast be approaching a reckoning for which its rash and reckless ingenuity has not prepared it. Unmoored from any stable moral framework, our AI overlords may already agree with the Melancholy Dane that “there is nothing either good or bad, but thinking makes it so.”

The readiness is all. And with Chinese autocrat Xi Jinping and a heedless, demented, “high-IQ” American president leading the charge of the bright brigade into whatever heaven or hell awaits humanity with the unfettered unleashing of “super intelligence”, we’re not ready at all.