It’s comforting, knowing the future. Repeat it like a litany, tattoo it on your dominant hand. Carve it into your computer screen, next time the thing looks at you all cocky, carnivorous:

Either AI programs will be able to produce expressive-seeming content, or they will not.

A steadfast axiom. With such solid ground underfoot, we might even forge a bit further ahead with some speculation about which of these two outlooks is more likely. Numbers are hard though, so I’ll pass on guessing any probabilities. The more basic question is whether option one, AI getting that good, is even a realistic possibility.

And to that I’d say yeah. Most definitely.

A select few won’t need much convincing on this front. See again that eager Google engineer. For most of us though (especially the creative workers who really have some skin in this game) there are no doubt some major hurdles to be overcome before we’re remarking to our neighbors that the old laptop sure seems chipper lately. The ever-trusted gut instinct, for one, is unlikely to have it. And how can we disagree when “seeming” is rooted in personal intuition, utterly subjective, abdominal?

We don’t actually need to. While subjective reactions are fluid and somewhat persuadable, the argument here isn’t that AI skeptics should, for any purely intellectual reason, override their intuitions and get with the more fantastical program. Our assessments of “seeming” do ultimately come down to instinct, but the measure of those instincts – the way something does factually seem to us, the way it is perceived – is an empirical issue that won’t be decided by theoretical insights alone (though they will have some influence). The case for the possibility of expressive-seeming AI is rather that there’s a realistic way for our skeptical gut reactions to undergo a natural and instinctive evolution. For our cynicism about AI to resolve itself, perhaps even against our conscious wishes, as the tech progresses. Not as much as a concession to new evidence but as a simple matter of fact, a shift in perception more than opinion.

That isn’t to say such a transition would come easy. The foundation of the skeptical outlook is a strong one; at its center the fact that, while the Hard Problem would suggest that we ignore “process” completely in evaluating things like consciousness and expressivity, we realistically aren’t going to – theory can tell us we should forget about it, but that doesn’t mean we will. In other words, we can’t erase the influence of our process-knowledge on our product perceptions. However humanlike generative art might seem in a vacuum, the machine which crafted it will naturally come under our scrutiny.

And given our knowledge about the generative machine and its process – that it’s just mapping patterns, crunching numbers, solving equations and connecting coordinates, all procedures that don’t require any level of conscious communication – it’s no surprise that AI outputs, when we know their origin, don’t readily strike us with the same expressive character that imbues human work with its special value. How could any mathematical mapping have the same appeal as a purely organic effort, those unique attitudes of vivid imagination, personal exchange, emotional intimacy, and directed intention? The easiest answer is that it can’t, and we can’t impugn our intuitions for defaulting to the path of least resistance.

Then again, even the hardiest naysayers will occasionally admit to a guilty sense of wonder at just how humanlike, imaginative, emotional, AI products can appear. It’s safe to say most of us who’ve played around with GPT at all have shared that experience, but most of us also still can’t (or don’t want to) shake the instinct that the expressiveness we perceive is only an appearance, a cheap imitation. A sense of “seems-but-isn’t”, which is effectively equivalent to “doesn’t-seem”.

But why is that? Why don’t we all take after LaMDA guy? Part of our hesitation can be attributed to the influence of our general process-knowledge, but I suspect this knowledge is acting more as a localized blind spot, impeding our perception of one specific expressive quality in AI outputs: intention. After all, examining the human creative intent is what clued us into the primacy of expressiveness in the first place. It makes sense that the lack of a seeming-intention would preclude us from experiencing a sense of connection with the creative force behind a product – without intent we don’t see art, we see accident.

And since we attribute the capability for intention based on, pretty much, whether the actor in question seems outwardly and internally human (both, again, evaluations of process), hunks of wire and circuitry going “whirr-buzz” in some Silicon Valley basement might be unlikely to get the nod, no matter how incredible their outputs.

In support of this point, I’ll propose the following: that were a pigeon to perform its impression of a flying ace with your forehead as its target, you’d be much more likely to suspect malicious intent than if a tree dropped an ill-timed pinecone. Get a monkey involved and you’d have no doubt that the resulting stain on your honor was a consciously directed expression of personal animosity. Tree, bird, and monkey are all living beings, but we scale our impression of each tormentor’s unforgivable meanness on our sense of their anthropomorphism. For basically this reason, I think we’ll struggle to call a mess of computer-stuff sentient or even expressive until it comes equipped with humanlike sensory systems, a more anthropomorphic internal structure, and maybe even a reassuringly bipedal android chassis, complete with rosy cheeks and fluttering eyelids – we’ve yet to resolve the ethical dilemmas involved with recognizing a full-fledged agency in children and animals, and robots are less naturally charming.

So by all indications, even if we should ignore the process AI uses in evaluating its products, we probably won’t. Maybe our unfair biases could be avoided by not revealing the provenance of AI generated content to the public, in a kind of Turing Test oriented toward expressivity instead of intelligence. And sure, again, maybe if the content passed that test, we should think of it as expressive. But once the treachery – however scientifically justified – came to light, we likely wouldn’t. Keeping the secret forever would just be cheating.

But that’s not the end of the story. AI’s outputs one day seeming-imaginative, -emotional, and so on – all the more superficial qualities of a product’s appearance, as it turns out – isn’t actually that difficult to envision. We’re already impressed on those fronts, and there’s no specific reason why it would be unrealistic to expect further improvements. Potential for programs getting that good at these aspects of expressiveness is clearly in the cards. Our hangups in imagining enough intentionality and anthropomorphism in the AI creative process, then, are the only persistent obstacles to conceding that a well-rounded expressive-intelligence might one day overcome our skeptical instincts.

So if we can find any realistic path, however thorny, for that intention and seeming-humanity to appear somewhere in the AI process, then a future where art machines get that-good is a genuine possibility.

Surprise surprise, there’s not much searching necessary. The path is already before us. And laid clear too, by virtue of the massive gaps in our understanding of neural nets and their progeny, lending that precious process-knowledge a certain Swiss-cheese quality. To review, the depth of our understanding depends on the level of description chosen – At the highest level, we have the machine’s directive: find patterns in the initial dataset, reproduce similar ones in the context of a given prompt. And at the lowest level, it’s all fancy math and hardware that we’re pretty sure could be explained by someone, somewhere, to whoever else has the right credentials.

But at the theoretical level, whether you place it highest or lowest, we’re essentially clueless. Whatever set of principles underlies the success of AI systems is an enigma. All we know is that we shaped them loosely in the image of neural connections in brains (there’s your anthropomorphism), set them up to model data and make good predictions, and that it worked – really, really well. So mysteriously well that researchers are regularly confronting “emergent abilities” in their programs which they’re completely unable to explain. These are, as the lingo suggests, capabilities acquired by AIs which weren’t programmed into them, but which materialized anyway, either out of basic existential belligerence or spite for the notion that humans are these great big smarty-pantses who tend to have a good handle on what we’re doing inside our fancy labs and research centers. A Google AI trained in English mysteriously picked up Bengali. ChatGPT can somehow run computer code and predict movie titles from strings of emojis. There is strong evidence that LLMs in general create internal models of the physical world – one of them arranged its own representation of the boardgame Othello not long after learning the rules.

Some analysts claim not to be impressed by these talents because however they emerge, they can ultimately be attributed to the program’s complex synthesis of its training data. Fine. But this doesn’t make the fact that they do emerge any less incredible, because we still have no clue why it happens, which makes further surprise impossible to exclude with any confidence. And there lies the avenue to seeming-intention – it would only be one more startling behavior to materialize suddenly from the frontier, another discovery made in the course of exploring the great generative unknowns.

Could a convincing expressive-intelligence then arise as an emergent property of AIs as they increase in size and complexity, even if it doesn’t originate from a real inner experience?

Yes. Take our incomplete knowledge of why generative AIs work, the fact that strange abilities are already issuing from them like freaks from the county fair, and frame these points in the context that the only real “theory” behind AI is that brainlike structures will do brainlike things: front and center in the picture we’re left with is a ready-made highway toward conceding a fully realized expressive-intelligence, in spite of ourselves. With the infrastructure already in place for us to shift away from our skeptical stances on an intellectual level, a paradigm shift on the instinctual level might simply require further improvements in generative outputs – the more the product quality increases, the more likely the path of intuitive least resistance will be to attribute that quality to humanness, intention, and expressivity. Even if we don’t really think they’re present – we can retain the view that the AI process is essentially un-humanlike, mathematical, pattern-centric, unconscious etc. and still concede that it might become humanlike enough, intentional enough, to win over our intuitive reactions to its outputs. Given strong enough products, it’s entirely feasible that all but the stubbornest skeptics could find themselves instinctively, empirically, conceding a seeming-intention, and therefore a seeming-expressiveness, to these pesky technological jumbles.

It appears everything’s really on the table, then.

So where does that leave us, we human creatives?

We’ve got a definitive prediction for the future – AI will get that good, or not. And we know that both of these scenarios are eminently possible, neither a mere hallucination.

Best be prepared for both, in that case.

But prepared for what? For battle? For submission? For a new era of creative flourishing, a constructive union of the human brainwave and the surging mechanical impulse? Or for the same old story – a technological panic followed by a whimpering withdrawal back into the status quo, maybe now furnished with more knobs and buttons.

Only one sort of question remains to those of us paint-covered, ink-stained daydreamers that have managed to retain any pragmatic sensibility as we pursue the odd magic of creation –

AI is upon us, treading the same path we are. Can we coexist? What should we do about it?