AI and Intention; or, What Comes After Contemporary Art
50+ years of "theory" are about to be undone by intentional machines.
We can take a recent high-profile conversation about “AI Art” — which, by the way, doesn’t exist, anymore than “Photo Art” or “Painting Art” or “NFT Art” does — between Jerry Saltz and David Wallace-Wells as indicative of how confused people in general probably are about artificial intelligence and what it means for art and artists today. It’s not that one of America’s most well-known art critics and one of its premiere journalists can’t muster a single salient point about what or how art means in the dawning age of artificial intelligence, it’s that the points they do make don’t stem from any coherent idea of what art is and how it might mean anything at all.
For example, an important point could have been made when Wallace-Wells queries Saltz about an image he “prompted AI to generate” based on a picture of a dog that is the subject of one of Francisco de Goya’s late “Black Paintings.” About the image that the AI generated, Saltz says that it “came from nowhere and exists in no time,” in other words, that is has no “context,” to which Wallace-Wells justifiably responds: “Isn’t the context that Jerry Saltz in 2025 was trying to make sense of the world of AI and what it was doing to visual culture?” To which Saltz replies, curiously enough, that he “didn’t intend this” image, and that, because he didn’t intend it, the image (which Saltz nevertheless calls “my image”) “arrives like the Virgin Birth.”
Saltz is obviously confused, while Wallace-Wells is obviously right, insofar as he offers a plausible account of what Saltz’s image is about — i.e. what it means. When Saltz says that he did not “intend” this image, we can take him to mean that he did not intend “exactly” this image, in the way that we can take Goya to have intended “exactly” everything that we see in his painting The Drowning Dog (1820-1823), upon which Saltz’s image is based. And there is a sense in which he is right. That’s not how AI image generation works. One (currently) puts in a prompt, and out comes an image. But in another sense, this point is irrelevant, because that’s not how any image generation — either mechanical or computational — works.
The point that Saltz is raising here, unwittingly one can presume (though not unintentionally), is a point that has been raised countless times with regard to photography, namely, in what way the things that photographs show us can be counted as intended by their makers and thus be considered to have a meaning (an answer to “Why” it’s in the picture). That Saltz doesn’t have a problem with photography as art can at least be assumed by the number of times he has written about it in just that context as an art critic.
But Saltz, and Wallace-Wells it turns out, are confused in a more fundamental way. While both writers appear to initially endorse some form of intentionalism as important for understanding what art means, or why and how it means what it might, both step back from that intentionalism just as quickly as they arrive at it. Wallace-Wells explains, with regard to Saltz’s AI-generated image, “Yet when I look at this image in particular, I don’t think I need to fold in the context of its creation. I just think, Those eyes are crazy!” To which Saltz responds: “I would love to see this work in person. I’d love to know what materials went into making it. I’d love to know its scale. I’d love to know its surface. I’d love to know what was in back of it.” Here Wallace-Wells responds to something that he sees in the image — the depicted dog’s “crazy” eyes — and Saltz responds to what he can’t see, which is the image itself — the materiality that this image doesn’t have but Goya’s painting does and that he imagines “his” picture could have.
Though it may be tempting to understand what one writer sees and the other doesn’t as significant, this is really a distinction without a difference, because in the “context” of what the picture means, either the dog’s crazy eyes and the imagined material aspects of this picture, including “what was in back of it,” can be considered germane to what the picture means or they can’t, insofar as what the picture means is either what was intended by the author of the picture, or it is whatever the viewer takes it to mean. With these comments, both writers appear to exit the intentionalist camp, though it is Saltz’s position that should be the more worrying: If Wallace-Wells doesn’t feel the “need to fold in the context” of the image’s creation (he can just focus on how “those eyes are crazy”), then Saltz doesn’t feel the need to account for the image at all; it’s not even art.
Much of this confusion — and it’s not Saltz’s alone; I take it to be a reigning confusion of our times (this is what makes Saltz the preeminent populist art critic of our day) — stems from two conflicts that are central to our understanding of visual art and which AI is bringing to the foreground in, if not new, then newly urgent ways. The first conflict is between “art” and “image” and what the status and relationship of these two terms has been and will be going forward. The second conflict is between two models of meaning: intentionalist and anti-intentionalist (or really, a model of meaning and one that isn’t). And though these conflicts appear distinct, they are intimately related.
Art and Image
There have been a number of attempts to theorize, or perhaps re-theorize, the “image” at a time when the proliferation of social and networked media has made images, both moving and still, something like the lingua franca of whatever it is that makes the “contemporary” what it is. Any brief survey of such efforts would have to begin with, or at least name check, W. J. T. Mitchell, who for 50 years has been writing about images in terms of what he initially called “Iconology” and then updated to “Picture Theory” and most recently dubbed “Image Science,” though it seems once images became networked and social, Mitchell’s preferred metaphors of the “clone” and the “fossil” (along with those of the “glove” and the “wrench” as figures for how images get “used”) belie a sensibility that isn’t quite equipped to handle the status of the image today.1
Peter Szendy’s The Supermarket of the Visible: Toward a General Economy of Images (2019) and Terry Smith’s Iconomy (2022) offer more recent attempts to account for images in terms of their profusions, their aggregations and flows, their currencies and circulations. The economic metaphors abound here, because the target is political economy in particular. The idea of an “iconomy,” central to both Szendy and Smith’s thinking, can be traced back to the period 1989-1992, which would give it some coincidence with the efflorescence of — take your pick — neoliberalism, the end of history, or the information age.
But the real point of theorizing the image in these terms is that it licenses the continuation of an anti-intentionalist theory of interpretation that has been a staple of theoretical discourse in cultural history and criticism ever since the late 1960s. What began then as a liberation of the “text” and of ecriture and of “discourse” itself from the agency of its authors — a liberation that was considered congruent with the emancipatory politics of the 1960s — has been translated into those same claims and the same politics now being made for images.2
As Szendy puts it towards the beginning of For an Ecology of Images (2021), his follow up to The Supermarket of the Visible (because “ecology” signals the compensatory moral ground one must occupy after visits to the “supermarket”): “I’m exploring here the possibility of an iconocentric ecology of images in order to open up an anthropofugal exit, to sketch out that step beyond a restricted economy of images produced by and for humans” (30), which he calls elsewhere a “nonhuman iconomy” (31). At a moment when a vast majority of images are produced only to be computed, not seen, whatever such images might be taken to mean can and must follow only from their effects and functions — i.e. what they do in the world.
The art historian and theorist David Joselit is perhaps the most significant avatar of this position. As he writes in (and about) his short programmatic book After Art (2012):
images possess vast power through their capacity for replication, remediation, and dissemination at variable velocities. In order to exploit this power for progressive ends, it is necessary to understand the potency of images on their own terms. To this purpose, After Art will shift critical emphasis from art’s production (and the corollary of artistic intention) to what images do once they enter circulation in heterogeneous networks (ix).
Joselit’s slip from “art” to “image” is narrated as a strategic shift from “object-based aesthetics” to a “network aesthetics,” a shift that entails what he imagines to be “a corresponding revision of critical methodology” (41) that would replace practices of interpretation — what an image, or a work of art for that matter, means — with accounts of how images circulate, “their insertion in networks where they are characterized by motion, either potential or actual, and are capable of changing format—of experiencing cascading chains of relocation and remediation” (12). The presumed politics for such a move are made clear when Joselit states that “assigning a meaning” to a work of art “is merely another way of setting an artwork’s price in the currency of knowledge, transforming it into a certain kind of commodity for collectors to buy and for museums to ‘sell’ to their audiences” (44). Meaning, in other words, is just marketing.
What is ultimately at issue in accounts such as Szendy’s and Joselit’s is the question of agency — who and what can claim it — and thus of intention — why one means what one does. As Joselit argues again in After Art, in a section on “Power,” prior to the modern era (whether we want to date this to the Renaissance or the revolutions of the 18th Century) the world was dominated by a visual culture of icons, pictures of divinities which possessed “their own power and agency” as objects of reverence and devotion, images that found use in ritual and sacrament. In the modern period, which the historian Hans Belting has called the “era of art,” Joselit writes that “such animate images ceded their agency to the artists who made them” (85). And thus, in the modern period, art becomes a vehicle for artists to explore that agency, as a means of both self-discovery and self-determination as autonomous subjects, especially as subjects in and of a disenchanted world in which one’s autonomy — i.e. one’s freedom — requires continuous efforts of authorization.
Today, Joselit and others hold that this “era of art” itself has come to an end. “Under the conditions of ubiquitous image saturation,” Joselit writes, modern art has “lost its urgency since everyone who inhabits contemporary visual culture assumes the complex communicative capacity of images to be self-evident” (88). After modern art, artists lose agency to the network, to the multi-nodal “formatting and reformatting of images” (88) that renders irrelevant the program of asking what artist’s intend for their work. Once that work is set free into the world, it is just one image among many subject to the whims of the web.
Is there agency in this schema? Is it back with the images, as in the era before the era of art? Or is it somewhere or sometime else? One answer, offered often enough, is that it lies with the “platforms,” the strata of technical protocols, interfaces, and infrastructures that compose the planetary computational stack. The scale and complexity of our contemporary era is such that any claim to self-determination or autonomy, let alone “freedom,” is regarded as laughable when our actions are circumscribed and wholly overdetermined by suprahuman forces, often of our own creation, whether these be governmental power, technological advance, or capitalist imperative. But this argument has been with us throughout the modern era; indeed one could go so far as to say that, as an extension of what Paul Ricoer called the “hermeneutics of suspicion,” it is constitutive of it. The metaphors may have changed — now we are to deal with formats and networks and “Iconomies” — but the strategic displacement of human agency, and thus of the artist’s intentions, and thus of meaning, remains the same.
There is however one agency that isn’t displaced and that is the theorist’s. When faced with the profusion and complexity of images, in their multiple formats, circulating through multiple networks, it is the adept theorist who arrives on the scene to analyze the flows and to forecast the trends. The theorist, once disposed to interpret, becomes the analyst of image “circulation” and “reverberations” (Joselit’s terms), the one who measures their impacts and effects and reports back what they, alone, can see.3 If in the era before the era of art, it was the priests and shamans who mediated the “power” of icons, putting them to use in the advancement of the faith; in the era after the era of art, it is the theorist, now recast in the guise of the data analyst, the trend forecaster or — even less sexy, but perhaps more powerful — the network administrator, who uses their access to advance themselves.
AI and Intention
That may sound unfair, but it’s not untrue, structurally at least. Because if the agency of the artist, what the artist intends by her work, is displaced or discounted, then it is only the viewers of that work — what they see in it, how they choose to understand it, the accounts they give of it, their “narrations” of its effects (on them; “those eyes are crazy”) or impacts (on others) — who can and will be enlisted in any account of what the work is about. In this there can be no disagreement over what a work of art means, because there is no question (no one asks) what the artist meant.4 The is the world that 50 years of anti-intentionalism has wrought.
Remarkably, today most artists are just fine with that, because most artists have been schooled in an academic system that, after the 1960s, largely accepted and promulgated theory-based arguments “against interpretation” (Sontag), ones which held that the significance of the work of art must remain “open” (Eco). These were arguments that celebrated the “birth of the reader” (Barthes) and dismissively asked “What matter who’s speaking?” when “discourse” can “circulate without any need for an author” (Foucault). After Duchamp demonstrated that any object could be art, and Warhol demonstrated that any image could be art, and Beuys claimed that anyone could be an artist, artists themselves have been resigned to let their audiences see in their works whatever they were going to see and so have given up on commitments to what they have meant.
As we can see with theorists such as Szendy and Joselit, soon enough the work of art, the work of the artist, doesn’t mean anything, except insofar as it is the proximate cause of certain effects, of “reverberations,” that the audience can hear but only theorists (like “operators” in the Matrix) can decode. Such an aggrandizement is also a deflation, however, for now one can see why the one-time authority of critics — their presumed place at the center of public forums on the “culture” — has waned over the course of more than a half-century during which the artist’s intentions — what her work is about, the things she does, which it was once the critic’s job to understand — have been so thoroughly dismissed. But the deflation doesn’t stop there. Soon after comes the rise and fall of the globe-trotting and institutionally-unbound curator, whose aggregations of artists and curatorial conceits are taken to confer value on the work they select. And in the wake of this super-curator’s demise, we now have the algorithmically-promoted and platform-bound influencer.
So whether we’re talking about some time after the end of art, or after the era of art, or just after art itself, it is still the same time, the time of beholders; the terms may have changed (from “art” and “objects” and “institutions” to “images” and “networks” and “platforms”) but the times are the same.
Or are they? What is all the more remarkable is that we are at the dawn of a moment with the advent of actually existing AI, a moment in which we may come to witness the displacement of human agency with the instantiation of machine intention — actions taken by AI unprompted by humans; actions which will have to be (indeed are being, now, today) interpreted in order to be understood.5
And this will actualize, in practice, what has been argued for for fifty years, in theory, but in reverse. Which is to say, for fifty years, theorists have imagined that things in the world — works of art once; images now — can have their own agency and so can have meanings that are separable from the intentions of the authors who made them; today there are (or very soon there will be) things in the world — AIs — that have intentions and will express those intentions by doing things, the meanings of which — the answer to the question of why they did those things — will be the condition of interpretation.6
There may come a point, as Ted Chiang imagined in his short story “The Evolution of Human Science,” in which our own capacity to interpret what the AIs are doing is exceeded. The authors of AI-2027, an influential forecast of the next two years in AI R&D, mark this as the advent of “neuralese,” an AI-specific communicative and computational “language,” and suggest that by 2027 there could be models of AI using advanced iterations of neuralese that even prior models of AI will not understand.7 As Chiang predicted, and the authors of AI-2027 affirm, this would leave us humans simply doing “hermeneutics.”
But this is just what a great many theorists of art and images today have chosen to do with the art of the past fifty years: create sophisticated, elaborate and yet imaginary accounts — imaginary because they cannot be disproven or disagreed with — of what things like works of art, or networks of images, do in the world, rather than accounting for what they mean because they are what the artist or author who made them meant to do.
In “Against Theory,” all the way back in 1982, when Walter Benn Michaels and Steven Knapp first challenged theory’s attempts to “split apart terms that are in fact inseparable” — i.e. “authorial intention and the meaning of texts” — they asked “Can computers speak?” The point of the question was to point out that, if we mean to answer the question, then “the only real issue is whether computers are capable of intentions,” and “however that issue is decided,” Michaels and Knapp argued, will rest on a “judgement as to whether computers can be intentional agents” (729). They acknowledge that “a great deal — morally, legally, and politically — might depend on such judgements,” but those moral, legal, and political dependencies will have no bearing on whether the AIs, as agents with intentions, must mean what they say.8
If we are going to take seriously the question of AI and Art and not repeat the mistakes and confusions of the past fifty years — mistakes and confusions that Jerry Saltz and David Wallace-Wells illuminate almost perfectly in their recent conversation about that non-thing “AI Art” — the judgement we need concern ourselves with is the one that Michaels and Knapp identified in 1982, and even reiterated just a couple of years ago in response to the rise of LLMs and a resurgence of interest in the claims the two made in “Against Theory,” and that is whether AI’s can be intentional agents.
If we judge that they can be — and by all accounts currently, with AIs that will have the capacity for dishonesty and so for sincerity (and thus irony), they can be — then it will follow that they can make art — not “AI art,” but “Art.” About which we can ask what that art means, because we can ask what the AI intended, which will be the same as asking what it meant, which is what we have been, or should have been, asking of artists all along.
With the return of intention, because of the advent of machine intention, perhaps we will finally find ourselves “after theory.” And whatever comes after that, it will be something new.
See Mitchell’s Iconology: Image, Text, Ideology (1987); Picture Theory: Essays on Verbal and Visual Representation (1995); What Do Pictures Want? The Lives and Loves of Images (2006); and Image Science: Iconology, Visual Culture, and Media Aesthetics (2015). LINK
As will be obvious by now to anyone who has read my work on this and other topics, I am following (as best as possible) the case for intention as integral to meaning, which has been made consistently, since 1982 at least (“Meaning is just another name for expressed intention”), and, of late, even insistently, by Walter Benn Michaels (and, in the case of 1982, Michaels with Steven Knapp). For how this intentionalism is applied directly to visual art, and its implications for politics (not emancipatory, but redistributive), see Michaels’s The Beauty of a Social Problem: Photography, Autonomy, Economy (2015).
In his most recent work, Art’s Properties (2024), Joselit will describe this agency in terms of “witnessing,” which doesn’t give an account of what a work of art, or even an image, might mean, it only “narrates” that to which the image already attests. As witnesses we don’t interpret; we testify. “To bear witness,” Joselit argues, “requires recounting one’s experience of an event. […] But the mechanics of narration are inseparable from the ethics of a particular telling, and consequently, any artwork can be assessed only through the contingent judgments of each person who sees it. In other words, no narration can claim to be definitive…” (104-5). However, the “ethics of a particular telling” are tied to the “judgement” of whether that telling is “responsible to its subject,” and yet, “there will be many such judgements as well as many principles upon which to define responsibility.” In other words, its narration all the way down. In this Joselit’s account is rendered even more paradoxical: either artists intentionally use the power of images to affect changes (in his words, a “fold, disruption, or event”) within the network of images, and thus the meaning of their work is just the actions that they take, or those actions can’t be counted as theirs, and thus it’s not clear that they, as artists, can be said to do anything at all.
Again this case is made by Michaels (and Knapp), and I have yet to read a retort of the very many that have been written that makes a convincing case otherwise. What is even more interesting is that, while the partisans of intentionalism — many of whom are affiliated with and are published at nonsite.org — regularly address and critique the anti-intentionlist camp, that camp — e.g. Joselit and others who have been affiliated with and publish in October — does not, as far as I have read, engage with or debate the intentionalist position.
The “AI 2027” report by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean is highly instructive in this respect and is required reading for anyone curious about the near future of AI. By later this year, the authors forecast, that the big AI companies will be using AIs to train other AIs, at which points the most advanced models can be expected to have “memorized” their “Specs” (“a written document describing the goals, rules, principles, etc. that are supposed to guide the model’s behavior”) and so will “learn to reason carefully about [their] maxims,” after which “the AI will hopefully be helpful (obey instructions), harmless (refuse to help with scams, bomb-making, and other dangerous activities) and honest (resist the temptation to get better ratings from gullible humans by hallucinating citations or faking task completion).” At this point, the author’s write, the AI companies’ alignment teams, should be asking the following questions: “Does the fully-trained model have some kind of robust commitment to always being honest? Or will this fall apart in some future situation, e.g. because it’s learned honesty as an instrumental goal instead of a terminal goal? Or has it just learned to be honest about the sorts of things the evaluation process can check? Could it be lying to itself sometimes, as humans do?” The authors then conclude that, “answer[s] to these questions would require mechanistic interpretability—essentially the ability to look at an AI’s internals and read its mind. Alas, interpretability techniques are not yet [in late 2025] advanced enough for this.” It stands to reason, however, that, though interpretability is the name of the game here, “reading” the AI’s mind will not provide us with answers to these questions.
See CEO of Anthropic Dario Amodei’s recent (April 2025) post “The Urgency of Interpretability.”
“We call this “neuralese” because unlike English words, these high-dimensional vectors are likely quite difficult for humans to interpret. In the past, researchers could get a good idea what LLMs were thinking simply by reading its chain of thought. Now researchers have to ask the model to translate and summarize its thoughts or puzzle over the neuralese with their limited interpretability tools.”
A nod here to Stanley Cavell’s Must we mean what we say? (1969).
About intentionality, it can't be intentionality alone.
There is this paragraph from Schopenhauer,
"Genuine works bearing immortal life arise only from such immediate apprehension. Just because the Idea is and remains perceptive, the artist is not conscious in abstracto of the intention and aim of his work. Not a concept but an Idea is present in his mind; hence he cannot give an account of his actions. He works, as people say, from mere feeling and unconsciously, indeed instinctively.”
Also, Schopenhauer distinguishes between Idea and Concept. Concept alone is sterile.