I once read a gallery press release for a show of abstract ceramics that opened with the phrase “the artist interrogates the porous boundaries between the body and its material residues.” Three hundred words followed. They cited Deleuze. They invoked “haptic perception.” They described the work as “refusing the teleological impulse of the vessel.” Nowhere did they mention that the ceramics were, by the artist’s own admission, the result of an afternoon spent making lopsided bowls and deciding they looked sad enough to be art. The press release wasn’t wrong, exactly. It was something worse. It was interchangeable. You could have swapped the artist’s name, the medium, the gallery, and the syntax would have held. A template wearing a theory haircut.
This is where the art world’s current panic about AI-generated text gets funny. For roughly two decades, institutional art language has been running on template logic. The vocabulary is finite. The sentence structures are predictable. The citations rotate through the same dozen names. If you fed every gallery press release from the last fifteen years into a training model, the output would be indistinguishable from what the art world already produces by hand. The anxiety isn’t that machines will write like curators. The anxiety is that curators were already writing like machines, and now everyone can see the machinery.
The Syntax Was Already a Closed System
Consider the verbs. A typical exhibition statement from a mid-tier gallery in any major city will contain at least three of the following: interrogates, subverts, engages with, reckons with, grapples with, probes, troubles, unpacks, situates, destabilizes. These verbs are never followed by anything concrete. The work interrogates—but what it interrogates is always “notions,” “constructs,” “boundaries,” or “the canon.” The canon is always being disrupted, never specified. The boundaries are always porous, never named. The work is always in conversation with something, but the conversation is always one-sided, and the other party is usually dead.
The noun phrases are similarly constrained. Site-specific means installed in a room. Research-based practice means the artist read something. Process-oriented means the artist would prefer you not judge the finished object. Multidisciplinary means unfocused. Emerging means under thirty-five. Established means over forty-five with gallery representation. Post-medium means the artist could not decide. Post-studio means the artist cannot afford one. The entire lexicon functions as a compression algorithm: it reduces complex aesthetic decisions to a set of tokens that signal seriousness without requiring thought.
Then there are the citations. Deleuze appears in roughly forty percent of press releases I have read in the last five years, usually invoked as a kind of theoretical patron saint for whatever the work happens to be doing. Bakhtin shows up whenever the word dialogic is deployed—which, in a recent three-month survey of London gallery statements, appeared in three unrelated exhibitions covering photography, installation, and, yes, ceramics. Rancière arrives when the work has political aspirations but no political content. Barthes is cited when the artist wants to declare the author dead while simultaneously signing the certificate of authenticity. These references function not as engagement but as credential. The academic equivalent of a verified checkmark.
Who Taught the Machine to Speak
The art world likes to treat its institutional language as if it emerged organically from decades of critical discourse. It did not. It was taught. And the primary teachers were MFA programs.
The MFA workshop, particularly in the United States and United Kingdom, developed a specific pedagogical syntax over the last thirty years. The critique format—where a group of students discuss one student’s work for an hour while a faculty member steers the conversation—rewards a particular kind of language. Vagueness is prized because specificity risks being wrong. Theoretical citation is rewarded because it demonstrates investment in the discourse. Emotional directness is penalized because it is “unsophisticated.” Over time, students internalize this and produce writing about their own work that sounds exactly like the writing they were trained to admire: dense, hedging, citation-heavy, and almost entirely empty of concrete description.
Museums adopted this language for a different reason: legitimacy. A regional museum programming contemporary work needed to signal that it was serious, that it belonged to the conversation happening in biennials and art schools. The fastest way to do that was to adopt the syntax. The wall text grew longer. The vocabulary grew denser. The Deleuze citations multiplied. The museum was not engaging with theory; it was wearing theory’s jacket to a party it was not sure it had been invited to.
Galleries had the most cynical motivation. A gallery press release serves one primary function: to make the work seem like more than it is, so that the price seems like less than it is. The institutional syntax is perfect for this because it is unfalsifiable. You cannot argue with a press release that says the work “troubles the boundary between presence and absence” because the sentence means nothing and therefore cannot be wrong. The language is a shield against the question that every collector eventually asks: is this worth what you are charging? The press release answers by changing the subject to Deleuze.
The Exposure Hypocrisy
What makes the art world’s response to AI text generation specifically hypocritical is that the same institutions now wringing their hands about algorithmic writing have spent twenty years producing algorithmic writing of their own. The panic is not about quality. It is about exposure.
The Authors Guild, in its AI Best Practices for Authors, frames the concern as one of preserving human voice: “As a writer, it is your original voice, thinking, and creativity that make you the writer that you are.” The Guild describes AI outputs as “generic mashups of pre-existing works ingested during training.” This is a reasonable position for novelists, essayists, and poets. It is a less reasonable position for the assistant curator who last year wrote a press release that began “the artist’s practice engages with notions of temporality, embodiment, and the residue of the everyday” and is now concerned that AI might debase the craft. The craft was already debased. The mashup was already happening. It was just happening by hand, one press release at a time, in offices from Chelsea to Berlin.
The deeper problem the Guild identifies—that AI threatens to make “quality human writing” a “rare luxury good”—already describes the art world’s relationship to its own institutional language. Quality writing about art is already rare. It is already a luxury. It exists in a handful of publications, on a handful of personal blogs, and in the occasional catalogue essay written by someone who actually looked at the work. The rest is template. The rest is the output of a system that learned to generate plausible-sounding sentences about art without requiring anyone to have a thought about art.
This is why the broader cultural context matters. Pew Research Center’s data on news habits and media consumption shows that fifty-seven percent of American adults express low confidence in journalists to act in the public’s best interest, and that Americans report feeling “worn out” by the very information ecosystem they say is essential to civic life. The parallel to the art world is not exact, but it is structural. Audiences across cultural sectors have been quietly withdrawing from institutional language—not because they do not care about the content, but because the language through which that content reaches them has become hollow. People sense, even when they cannot articulate it, that they are being addressed by a system rather than a person. The art world’s press release is the cultural equivalent of the automated email that says “we value your inquiry.” A sentence that means the opposite of what it says, generated by a process that has no stake in whether you believe it.
What a Template Detector Would Reveal
If you wanted to test the claim that art-world language was already algorithmic, the experiment is straightforward. Take one hundred gallery press releases from the last five years. Strip the artist names, gallery names, and medium-specific details. You will find that the remaining text falls into approximately six recurring structures: the ontological inquiry (the work “questions what it means to…”), the political gesture (the work “responds to the conditions of…”), the material investigation (the work “explores the limits of…”), the spatial intervention (the work “reconfigures the relationship between object and viewer”), the temporal meditation (the work “slows down the act of looking”), and the institutional critique (the work “exposes the structures of display that…”). Within each structure, the vocabulary rotates through the same thirty words. The citations cycle through the same fifteen theorists. The adjectives are drawn from a pool of perhaps fifty approved terms, all signaling seriousness and none requiring specificity.
The same test could be run on curatorial statements for group exhibitions, which are even more constrained because they must describe multiple practices under a single thematic umbrella. The thematic umbrella is almost always a single word drawn from critical theory—trace, residue, threshold, encounter, friction, entropy, repair—paired with a vague noun: the trace of the everyday, the threshold of the visible, the friction of the social. These titles function as passwords. They signal that the curator has read the right things and expects you to have read them too. They communicate nothing else.
When an institution’s communications workflow has devolved into filling in a stable template with interchangeable vocabulary, the line between human and machine authorship becomes philosophical rather than practical. A curator using an AI script writing tool to draft an exhibition statement is not introducing automation into a previously manual process. They are making explicit the automation that was already implicit in the institutional syntax. The tool does not create the template. It reveals it. The output will be indistinguishable from the hand-written version because the hand-written version was already following the same algorithm—just more slowly, and with more self-importance about having done it.
The Real Fear
What the art world is actually afraid of is not that AI will produce bad writing. Bad writing has been the industry standard for decades. The fear is that AI will produce the same writing, instantly, for free, and without the social performance of a curatorial assistant staying up until 2 a.m. sweating over a Deleuze citation. The fear is that the labor of producing institutional language—the hours of pretending to think hard about sentences that were always going to come out the same way—will be exposed as unnecessary. If a machine can generate a press release that passes muster in thirty seconds, then the human who spent three days on it was not exercising craft. They were performing devotion to a system that never needed their devotion.
This is why the art world’s response to AI has been so disproportionately anxious relative to the actual threat. A painter is not threatened by AI text generation. A sculptor is not threatened. The people who are threatened are the ones whose job is to produce the verbal wrapper around other people’s visual work—and whose authority depends on the belief that this wrapper requires specialized human intelligence. AI reveals that it does not. It requires a template, a vocabulary list, and a willingness to sacrifice specificity for plausibility. These are things a machine can handle. They are also things a second-year intern can handle. The difference is that the intern will eventually learn to have opinions, and the machine will not. But the system was never designed to reward opinions. It was designed to reward compliance with the syntax.
What Honest Art Writing Would Look Like
The way out of this is not to ban AI or to police its use in institutional contexts. The way out is to stop pretending that the language the art world has been producing is worth protecting. If the entire apparatus of gallery press releases, curatorial statements, and wall texts can be generated by a machine trained on the existing corpus, then the existing corpus was the problem, not the machine.
Honest art writing would begin by describing what is in the room. It would name the materials. It would state the scale. It would admit when the work is confusing, ugly, dull, or derivative, and it would try to figure out whether those qualities are failures or points. It would cite theory when theory is actually doing work in the piece, not as decoration. It would say “the artist made lopsided bowls and decided they looked sad” if that is what happened, rather than constructing a sentence about the teleological impulse of the vessel. It would be willing to be wrong, which is the one thing a template can never be.
The art world will not do this voluntarily because specificity is risky and vagueness is safe. A press release that says “the work interrogates the body” cannot be contradicted. A press release that says “the work is a small ceramic bowl that appears to have been dropped” can. But the second sentence tells you something. The first one tells you nothing. The art world chose nothing, systematized nothing, taught nothing in MFA programs, institutionalized nothing in museums, and monetized nothing in galleries. Now a machine can produce nothing faster than they can. The appropriate response is not legislation. It is embarrassment.
The best thing that could happen to art writing is that AI forces a reckoning with the fact that most of it was never writing at all. It was formatting. The artists will survive. The work will survive. The press releases will survive, because someone has to write them, and if a machine writes them, the only thing lost is the illusion that they were ever worth reading. The real loss—the one nobody is mourning because nobody noticed it happen—was the slow replacement of critical voice with institutional syntax twenty years ago. That was the automation. AI is just the mirror.