The first of these issues is particularly complex and stubborn. Luckily, that means it’s seen the business end of enough serious scholarship to put Professor Owl in a coma. But less fortunately, most of the established theory points to the problem of machine-consciousness being unsolvable – the best answers give a resolute “maybe”, the kind of determinate shrug wielded most effectively by Zen monks, shameless criminals, and hopelessly bewildered academics.

Always tempting to simply abandon these kinds of impossible conundrums. Whether fully conscious tin-supremacist robots are going to culminate the apocalyptic lineage of spinning looms, conveyor belts, and computers, science can’t tell us either way. Still, the creative populous is interested mostly in a single feature of consciousness – expressivity – and making a more limited prediction on that front will at least help us decide whether we should begin constructing underground doomsday studios or can continue business as usual, plug away freely at our desks and knitting chairs and easels without listening over our shoulders for the plagiaristic whirring of smart-appliances.

That means we can’t ignore the more general consciousness question completely. To make any good predictions about whether machines could become expressive, we’ll have to know our limits in evaluating their capacity to develop or imitate an inner experience. Understanding why machine-consciousness is so tricky to assess will help us decide just how well we can play the role of psychic detective, what kinds of evidence we might find of an illicit artistry lurking within a neural-net’s electric webbing, and what we can write off as noise, uninterpretable static.

The Hard Problem of Consciousness

Our confusion as to the plausibility of machine-consciousness is rooted in a startling ignorance about the nature of consciousness in general. A bafflement dubbed, with some tongue-in-cheek optimism, the “Hard Problem of Consciousness.” Titled and formalized thusly in a 1994 speech by philosopher/cognitive scientist David Chalmers, who may have surmised that calling it the “Unresolvable Problem” wouldn’t win the thing much industrious attention, or philosophy departments any funding.

The Hard Problem is concerned with how, why, and whether cognitive processing entails a subjective experience. Chalmers first put it this way: we know the brain has functions that give rise to behaviors, but “Why is the performance of these functions accompanied by experience?” This dilemma is so transcendently stubborn because, unlike “easy” problems that seek empirical explanations (armchair brainteasers like discovering why different types of neural activity give rise to certain behaviors, how sensory systems process information, and “explaining the dynamics of access consciousness in terms of the functional or computational organization of the brain”) the Hard Problem seems to innately defy any type of objective investigation: if there isn’t any necessary reason why an inner experience must accompany either brain activity or behavior (the only two observable phenomena we associate with consciousness) then it’s impossible to study that experience empirically. In other words, the only consciousness anyone can access is their own; we have no means of directly interacting with – or even confirming the existence of – the subjective states of others.

The trouble is demonstrated best by a group of thought experiments sharing as a protagonist the redoubtable, mind-starved, “philosophical zombie.” In a nutshell, the zombie is physically and behaviorally indistinguishable from a human, leading a normal human life and even boasting a normal human brain with all the associated electrochemical activity. But it’s lacking something crucial: it feels none of the subjective states that characterize human consciousness, and so has no inner experiences of its own to speak of. Though it would speak of them anyway – scratch at mosquito bites without feeling itchy, mumble in its sleep without really dreaming, laugh at jokes without joy, weep when Bambi's mom dies without feeling any sadness.

Every scenario in which the zombie appears boils down to a single conundrum – it would be impossible to differentiate any standard issue human from a zombie impostor. So how do we know that our own peers and friends and family aren’t just zombies, vacantly going through the motions?

Beats me, man, say the experts, shrugging.

And so says the Hard Problem: you can’t tell if anyone or anything outside of your own personal self has any inner experience. I can’t tell if you do, you can’t tell if I do. And if we can’t even say for certain whether other humans are conscious, then of course we’re suspicious about the status of machines as they become, if not more humanlike on a fundamental level, at least more zombielike on a superficial one – seeming to exhibit qualitative thought processes and mental states that they never truly experience.

Hence the big ‘maybe’ to the possibility of machine consciousness, and the need to understand the uncertainty around it before we can make any judgements about whether an emerging machine-mind might develop an expressive capacity – with a rather strict limit on anyone’s power to recognize any consciousness at all outside of their own precious skull, we’ll need to modify our assessment of machines accordingly.

Seemingness and Practical Consciousness

Fact is though, we do have a working assessment of human consciousness that can be applied just as well elsewhere. Thought experiments are all well and good, but for all save the most enthusiastic paranoids, who aren’t good role models, it’s completely untenable to navigate daily life continuously questioning whether your friends and neighbors are all mindless zombies. In the end, we simply trust that other people are conscious just like we are.

Why? Because they seem like they are. And we have no better standard. For practical purposes, seeming-consciousness is in effect the real thing.

There is at least good precedent for this strategy. Aside from however many thousands of years of it working well enough for any human being who didn’t turn out chronically twitchy and elaborately adorned in tinfoil, it was introduced to the modern academic canon by Alan Turing in his groundbreaking 1950 paper Computing Machinery and Intelligence, as the basis of what he called the Imitation Game, now known more popularly as the Turing Test.

In brief, the test goes: an interrogator is tasked with interviewing two subjects using typed questions and answers, one a real human and one a machine impostor. The interrogator then decides which subject scrubs their armpits in the shower in the morning, and which lacks the blessings of an organic nature. If interrogators can’t accurately distinguish man from machine in most cases, the machine wins the game and passes the test – meaning, per Turing, that it could be described as possessing something akin to intelligence.

Turing was incredibly prescient in his predictions for the future of AI. Partly because he was a genius and partly because he’s one of the pioneers of machine learning; his ideas set the trajectory for developing neural nets and remain in their DNA. A machine “whose manner of operation cannot be satisfactorily described by its constructors” would have been no surprise to him, nor would the success of a “digital computer with a random element” at achieving seemingly impossible functions. Turing was, in fact, the first notable figure to ask explicitly “can machines think?” – this very question was the foundation of his seminal paper introducing the famous Test.

In wondering the same, though with different underlying motives and more definitive confirmation of the devastating aptitude of computers, we’ve stumbled into a thought process very similar to Turing’s. He deemed the original question “can machines think?” to be “too meaningless to deserve discussion” and proposed that a sufficient answer could instead be achieved by assessing the human perception of machine capability – an endorsement of seemingness as the principal measure of an otherwise invisible mind. His dismissal of the original phrasing was probably even motivated by, along with a desire to specify its clearly ambiguous language, an awareness of the yet-to-be-formalized Hard Problem – in defending the validity of his test, Turing admitted that “the only way to know that a man thinks is to be that particular man” and called the belief that everyone thinks only a “polite convention.” Hence the necessity of establishing another reasonable “convention” for assessing the thinking-status of machines. To this point, we follow Turing’s example near exactly.

Where we differ is only in our invocation of ‘consciousness’ where he prefers ‘thinking’. But this is only natural. Before the days when machines could whip up a pastel portrait of King Kong and Frankenstein square dancing in 60 seconds flat, the imagination was burdened enough by the prospect of computers possessing any kind of analog to human thought processing – with neural nets now in vogue, we’re freer to explicate a distinction between consciousness and mere cognition which rested in the background of Turing’s arguments.

The Limits of the Turing Test and New Standards of Seemingness

Therein, and much telegraphed, we’ll find the thrust of the seemingness with which to penetrate the possible futures of our artificial adversaries, hopefully fending off the scary ones. For decades, the Turing Test was the standard benchmark to determine the seeming-conscious of AI programs, at least in popular perception. Now multiple AIs have somewhat quietly passed that test, including ChatGPT and Google’s LaMDA, which famously convinced one jumpy engineer that it was a fully sentient being. But the reaction from experts has been nearly unanimous in denouncing these results as irrelevant to the actual ‘intelligence’ of the offending programs. Instead, they’re used as evidence that the Turing Test is an insufficient measure of mechanized intellect, showing only that the winners of the Imitation Game are, well, good imitators.

The goalposts haven’t been moved without reason – I think most would agree that neural nets are engaged in something like ‘cognition’, if we understand that term as applying to any information-processing analogous to the kind that occurs in the human brain – again, the whole mysterious success of neural nets is predicated on the fact that they model this type of processing. Passing the Turing Test might be enough to describe such programs as ‘thinking’ in that weak sense, and especially would have been in 1950 when this result would have been nearly unfathomable.

Now that the horizons for AI tech are vaster, scientists are simply less easily impressed. We now have higher standards for seemingness, which is, unavoidably, a subjective measure contingent on our own fluid perceptions. Focus has more recently shifted to a search for artificial general intelligence (AGI), which might be described as a machine ‘thinking’ successfully about a wide variety of tasks/subjects rather than a singular function (like tricking an interrogator in a chatroom). Experts are devising more rigorous assessments, like the AI Classification Framework (ACF), that can evaluate for this more substantial kind of machine cognition. The ACF, for instance, tests for different categories of intelligence which include the classic linguistic-verbal and logical-mathematical intelligences as well as more exotic varieties like “existential” “musical-rhythmic” and “bodily-kinesthetic” intelligence. So far no AI has performed very well in these capacities.

Cognition vs. Expression: What Creatives Actually Care About

What’s happened here is that science has pretty much abandoned any explicit investigation of the possibility that a machine could seem-conscious, in favor of more concrete measures of seeming cognitively capable. It might be tacitly assumed that a program with impressive enough marks on the AGI front would also exhibit the quality of seeming-conscious, but not much dedicated effort is being put to investigating that quality for its own sake. This approach is all well and good for the scientists, since it avoids the Hard Problem and establishes AI research on solider empirical grounds: it eliminates the need to evaluate whether AI is undergoing a conscious problem-solving process to complete a given task, instead measuring the degree to which machines do, factually, complete the kinds of tasks that we consider cognitive.

But creatives need to make a different distinction, analogous to the consciousness/cognition one now being employed in the sciences, but more targeted on our unique situation. We don’t care how well a machine can do math or navigate an obstacle course. In fact, we don’t even care if it can paint an incredible still-life or write an “expert-level” essay or musical composition. We care if those all those technical achievements also constitute expressions. The relevant distinction for creatives, then, is between experience – whether in the process of generating outputs, AIs undergo subjective states of emotion, imagination, intention etc. – and expression – whether the AI-generated products seem to be imaginative, emotional, communicative, and so on.

It’s only when machines appear to have developed an expressive-intelligence that the human touch will be vulnerable to automation.

And thanks to the Hard Problem, we know that the seeming part of this equation is the only one with a practical application. In the end we won’t be able to guess at the existence of inner subjective states or expressive intention in future neural nets; if they ever do develop an inner experience, it will remain unavailable to us. Understanding the generative process might inform our personal decisions about using the tech, but it’s the products of AI, not any quality of the process it follows, that should be the focus of our projections for its future.

The Prophecy: Could Machines Become Expressive?

Which means we’re now equipped with all the necessary material to make an act of prophecy worth our while – we need to make a prediction as to whether an expressive-intelligence could arise in generative machines, even if it doesn’t originate from a real inner experience. If we play our cards right, I think we can rival even the finest of Turing’s own auguries, read with such impossible accuracy from the wiry entrails of AI’s earliest ancestors.

Arrogance? Never – Turing had only his genius to stand on; we’re positioned a good 70 years closer to the future, and a better vantage point – plus, sure, a confident delivery – is the key to all successful divination.