The cinematograph, the “noematograph,” and the future of AI art

You’re in Paris on the evening of December 28, 1895. It’s cold and windy, and this lull in the year, between Christmas and New Year’s, feels restless and full of possibilities. You are in the Boulevard des Capucines, a few blocks from the shore of the Seine and the Jardin des Tuileries. You stop in front of the Grand Café, a fashionable meeting spot for the elite.

Along with thirty-some Parisians, you pay a franc to the mustachioed gentleman at the door before descending into the Salon Indien, an opulent room in the basement decorated with Oriental motifs. Rows of chairs have been set up before a white screen, and there, propped up in the middle of the seats and aimed at the screen, is a strange, boxy machine with a hand crank and a lens.

This machine is the cinematograph, which promises to “capture, through a series of instantaneous images, all the movements that during a given time occurred in front of the lens, and then reproduce these movements by projecting their life-sized images onto a screen in front of an entire audience.”¹

Why have I taken you to the world’s first movie night? (Sorry, no stale popcorn or overpriced soda, not yet.) Because when you’re looking forward to the birth of a new artistic medium — in our case, AI — it may be helpful to look back at the birth of a mature one.² (Just to be clear, I’m not talking about AI as a tool that artists can use to create art in existing mediums — such as books, paintings, photographs, movies — but as a first-class techno-artistic medium in its own right, the equivalent of cinema, VR, and video games in the 20th century.)

AI as a promising medium for human artists may sound like an oxymoron. Indeed, surveying the reportage reveals a spate of failed attempts by AI to unseat humans, AI-generated product images that defy the laws of physics, and outright frauds. If AI truly is an emerging medium for artists, these are not auspicious beginnings.

Was the birth of cinema any more promising? On that December night, the inventors of the cinematograph, a pair of rather aptly named brothers, Auguste and Louis Lumière (along with their father, Antoine), showed their dark basement room audience a program of ten motion pictures. Shot on 35-mm film at 16 frames per second, the hand-cranked movies were each just under a minute in length (or about 17 meters, if you substitute space for time).

For the most part, these are scenes of everyday life captured with a fixed camera in one unbroken shot: working women in long dresses exiting a factory alongside the occasional dog, bicyclist, or carriage; serious men in hats walking off a boat, smoking and carrying canes and cameras;³ a baby held up by her father (one of the Lumière brothers) terrorizing goldfish in a bowl; bathers frolicking next to the sea, diving into the waves and emerging in an endless circle like some Escher drawing come to life; a horse-drawn bus rattling through the street. 

There is no editing, no special effects, no “art” as we’ve come to expect. With only these Lumière reels to go on, could any member of the audience that night have imagined The Godfather, Ran, Fight Club, Memento, Breaking Bad, Fargo, Vine, TikTok? Could any of them have foreseen that one day, telling stories through moving images would become by far the biggest component of our collective artistic and cultural output (whether your preferred metric is influence, number of consumers, or money)?

No, of course not. The beginning of a medium offers few clues as to what it will evolve into. Tim Berners-Lee couldn’t have foreseen in 1989 that his invention would one day host Pump.fun, ChatGPT, and YouTube. Likewise, the Lumières, who viewed their invention as a “scientific curiosity” and wanted to reserve it for scientific uses, refused to sell their combination camera-projector to anyone, not even to Georges Méliès, who offered them 10,000 francs.

“(A) horse-drawn cart moved toward us, followed by other vehicles and then pedestrians — in a word — street life!” Méliès exclaimed at the first motion pictures.⁴ “We stared flabbergasted at this sight, stupefied and surprised beyond all expression.”

Méliès would go on to make more than 500 films, including the cinematic milestone Le Voyage dans la Lune just seven years later, and play a big part in the invention of the language of cinema. What did this stage magician see in the cinematograph that the Lumières couldn’t?

Go and watch those films again. No, not on the phone while you’re riding the subway or between YouTube shorts, but in the dark, when everything is quiet around you, when your mind is free of distractions, and on as big a screen as possible. Let the moving images take over your visual field.

A century and a quarter ago, those women walked out of that factory, those men disembarked from the boat, and that baby plunged her hand into the fishbowl. They’ve all long since turned to dust, but here they are, right in front of you, laughing, moving, living. Now, do you understand?

That feeling down your spine? Wonder.

That was what Méliès experienced in the Salon Indien. The early Lumière films, academically termed “actualities,” weren’t so much a precursor to documentaries as a manifestation of the raw, foundational nature of cinema: Movement in time — presence, life, actuality — captured and then played back.

The film camera could take you to any place, any time, real or imagined. Whether you call cinema a time machine, a teleportation machine, or an immersion machine, at the core of film’s appeal is its ability to replicate, to simulate, to fabricate the experience of actuality, the feeling that you are there. Every trick in the language of cinema — the close-up, the tracking shot, the jump cut, the montage, the establishing shot, the L cut and J cut, the POV shot, the flashback, bullet time — is meant to enhance, manipulate, subvert, fracture, challenge, warp, and otherwise engage with our experience of actuality, of being there. This is how and why cinema could shatter the Aristotelian unities, resurrect orality in the age of literacy, and become the technological medium par excellence imitated by new mediums like video games and VR. 

Even in the primitive single shot of La Sortie de l’Usine Lumière à Lyon, one could already glimpse the magic of playing with actuality, the tiny seed that would eventually grow into the magnificent, towering sequoia of cinema. (And, of course, a stage magician like Méliès, who was already used to playing with actualities in his own medium, would see that potential before others.)

What is the equivalent of that seed in AI art?

I don’t have any certain answers, and I doubt anyone else does either. This is why I’m writing an essay.

Like most interesting questions, the best way to find an answer is through play. As a writer, I enjoy playing with words. Let’s apply AI to one of the oldest word-based arts, translation, and see what interesting results come out.⁵ This experiment feels particularly apt as the innovation that initiated the recent surge of AI capability in large language models (LLMs), the transformer, was initially developed as a translation technique.⁶

One of the commonest criticisms levied against human translators is that they have “agendas.” Instead of faithfully rendering the original text into a new language, like a camera faithfully reproducing onto film the light and shadows glimpsed through the lens, “bad” translators inject their biases into the text, polluting the reading experience. By that logic, a machine ought to be the best translator, for it is understood by many to be without “agenda” or intention.

I ask the machine⁷:

Give me a Classical Chinese poem that is relatively obscure to Western readers.

The machine suggests《登幽州臺歌》(“Song of Ascending Youzhou Terrace”) by 陳子昂 (Chen Zi’ang, 661-702 C.E.) from the Tang Dynasty:

Give me a translation.

Instead of critiquing the aesthetics of this translation (short version: It’s not good enough to even bother criticizing), I’ll focus on the “meaning” (however you want to define that amorphous concept). As someone who knows this poem quite well (it was one of the first Chinese poems I studied as a child), I’d say this translation is “right” but incomplete.

This translation feels odd to me. Can you try to give a rendition that is more in line with the way most contemporary Chinese readers would understand it?

Ah. Chen Zi’ang was a reformist scholar and poet who lived during the reign of Wu Zetian, perhaps the most famous of the many women who ruled the various Sinitic states in China’s long history. Chen had many policy differences with Wu Zetian and was frustrated that the sovereign ignored much of his advice. This political disappointment is evident in the second translation but not the first. Most critically, the second translation renders 古人 as “sages of the past” and 來者 as “enlightened leaders to come.” These translations are correct in the historical context of this poem,⁸ but if you were to look these words up in a dictionary, you’d get something much closer to “ancients” and “future ones.”

Which translation is more “agenda-free” or “transparent”? Time to get more playful.

Craft a few more interesting translations and tell me the agendas behind them.

Fascinating. I can’t imagine making these translations myself. They’re so divergent from one another (and my reading) that one can scarcely believe they are translations of the same text. And yet, after some reflection, I confess that they are all interpretations that could be extracted from the original.⁹

Are these translations, each driven by an explicit agenda, worse than the previous two?

If I had to pick, I would say that the “sages and enlightened rulers” interpretation, being closest to (what we know of) authorial intent and (what we choose to pay attention to in) historical context, is the most “accurate” translation, but that is because I, as an author, am especially protective of authorial intent. (As you can imagine, I’m sort of a pain to translate.) For those who find the presence of the author in the text a nuisance — either because they believe artists should “shut up and sing” or because they value their own interpretive freedom more — then fidelity to authorial intent is itself an agenda.

What about the first translation, the one closest to dictionary definitions out of this bunch?

Can you tell me the agenda behind the “literal” translation you did first?

The AI goes on, talking academically and authoritatively about the “priority on literalism,” “universality of emotion,” “avoidance of interpretation,” and “emphasis on solitude.” But the underlying point is clear: This translation of the poem is so divergent from the common understanding of Chinese readers because it adopts a set of choices about Chinese–English translation that detaches the text from its historical and cultural context and interpretive tradition in the source culture, and repositions the text as an independent artifact that can be interpreted by the rules encoded in the Western construction of “Chineseness.”

This is also an agenda.

Of the many possible translations derivable from a single text, the classicist Emily Wilson had this to say: “The real question for a translator isn’t do you want to tell the truth or do you want to lie … it’s which truths to tell. There are many true facts about any original text, not all of which could ever be conveyed in another language.”¹⁰

A translation only appears to be “without agenda” to a reader when the interpretive frameworks and cultural assumptions behind the translation align with the reader’s. Every translation requires a translator, a subjectivity. In the case of an AI-produced translation (at least the AIs based on the current, dominant architectures), the subjectivity is extracted/excavated/synthesized from the embedded subjectivities in all the linguistic examples it trained on.

All meaning-making requires subjectivity, not just translation. Every reader must first inundate a text with their subjectivity before meaning can be extracted, a truth we all are too aware of, surrounded as we are by nonstop controversies over interpretation in our symbol-saturated modernity. Just remind yourself how we cannot agree on an “agenda-free” understanding of the Second Amendment or how, if you have enough followers, any statement you post to social media will become contentious. (Every reading requires a reader and every meaning requires a meaner — sometimes a very mean one.)

Enough. Let’s get back to AI. What’s interesting here is that I was playing, alone, with a Chinese poem that had relatively few human-composed English translations. Nonetheless, ChatGPT conjured up a bunch of translations done by excavated/recombined/sythesized subjectivities, a (ghostly) sense of “someone stands behind these thoughts.” The question that matters here isn’t “Which translation is most accurate?” but “Taken together, do these translations make the poem more interesting?” And the answer to that is a resounding yes. 

The implied translators behind these divergent translations restored the kind of textural richness inherent in the original composition, which is lost in any single translation. I conversed with the subjectivities behind the words, probed them, interrogated them, and found myself with a new understanding: AI is a machine that captures and projects subjectivities, a “noematograph.”¹¹

To be sure, the captured subjectivities are crude and low-resolution, but already there is a thrill in my experience, similar to the thrill Méliès experienced when he first saw the machine that captures and projects actualities.

When I look back over the excitement surrounding AI-generated art, the same pattern emerges. People use AI to play with captured subjectivities — excavate, fragment, enhance, strip, recombine, edit, synthesize, and otherwise manipulate — and project the results back: A woman turns to her AI boyfriend, a responsive embodiment of cultural narratives about feminine desire, to co-author a fanfiction of their life together; enthusiasts one-up each another by pushing image generators to distill the scraped collective unconscious into “make it more” trends or to turn cultural stereotypes into a community picture-story featuring endless heists of the Mona Lisa; fans make and share chatbots embodying popular fictional characters with transformed personalities.

An LLM captures (and fractures) and projects back subjectivities as a cinematograph captures and projects back actualities. Just as manipulating actuality is the soul of cinema, playing with subjectivity is the soul of AI as an artistic medium. Whatever genres of AI art emerge in the future and however the language of AI art evolves, the seed (or at least a seed) will be this raw, foundational nature of the medium.

Much as the printing press changed the way we consume books, AI will change the way we consume text, music, and film. As the translation example shows, AI will help us probe the multiplicity of subjectivities inherent in every work of art and, in that process, better understand our own subjectivities.

But what about the art we can create with AI? What will that look like?

So long as we think of AI as just a way to replace humans, to do what we already do (but faster, cheaper, and more satisfying to our capitalist masters), it will produce nothing but trash.¹² AI-produced novels and images and movies will be no more interesting than filmed stage plays or photographs of paintings — a medium cannot succeed by imitating existing mediums; it must invent its own language, evolve its own genres, discover the kind of stories that only it can tell.

While it’s impossible, at this early moment in the metaphorical Salon Indien, for us to glimpse the full potential of the noematograph as a medium, there are tantalizing clues already on the screen. A notable feature of contemporary AI art experimentation is that while we’re often excited by the process, we’re just as often bored by the result.¹³ Despite all the headlines promising that AI will write better books than humans, I can’t find a single AI-written novel that has delighted readers. Not one. I enjoy playing with AI image generators, but I seldom go back to look at the images after a session. Asking ChatGPT to generate a bunch of translations to experience the Chen Zi’ang multiverse is interesting, but the actual translations, qua works of literary composition, are bland despite their explicit agendas. Talking to a chatbot is fun, but reading someone else’s chatbot transcript isn’t (except insofar as you learn tricks to use in your chats). Participating in an act of collective storytelling like “Mona Lisa: AI Cinematic Universe” is thrilling, but the final output is no more enjoyable to look at than the aftermath of an alternate reality game.¹⁴

There are multiple possible explanations for this essential “boringness” in AI-generated art. For one, current models just aren’t very good — this may change over time. For another, the default output mode for contemporary noematographs (LLMs and generative image models) is null subjectivity — an  inoffensive, diffuse, “this is what the internet thinks” sort of feckless, collective, no-voice (one further muffled, as always, by the AI company’s ass-covering rules). This averaging out of all the subjectivities the machine has been trained on  spews out cliches — which is great for research, when you are asking for facts and conventional beliefs. But for art to be interesting, it requires distinct, focused subjectivities, the drive to stake a position and to get away from cliches.

To get the AI to do anything interesting, the user must prompt it to abandon the null subjectivity and take on some specific stance (as I did in the translation examples above). The process of interacting with the AI to get it to take a position, to be someone, is where much of the fun of playing with AI lies.¹⁵

This suggests that AI, as an artistic medium based around playful inter-subjectivity, will be more about process than outcome, more about participation than consumption, more about the doing rather than the beholding. Some of the best AI art experiences are about what the AI prompts in you, rather than what you prompt the AI to do. Enjoying AI art will not be like reading a book or watching a movie, but more like having a meal with your friends or playing a board game. It is irreducibly ephemeral and personal, just as subjectivity itself is. It appeals to a person, not people.

An AI translation, for instance, may be nothing like a human-crafted translation. Instead of an “authoritative” but derivative literary text that encapsulates a single subjectivity’s experience of the original text, an AI translation may be more like an interactive experience through which the reader excavates her own reading by interrogating an infinite number of synthesized translator subjectivities to fully experience the potential inherent in the original. It will not replace a human translation; it will be something completely different.

Will the medium of AI develop genres that are best experienced alone (such as using AI to explore and excavate a work of art) as well as genres that are best experienced together (such as using AI to create a shared story that teaches you more about one another)? Perhaps this new medium, even more than games, will demand the consumer to be their own artist — a creasumer — not unlike dancing or playing an instrument for fun. It may render the very distinction between authorship and readership nonsensical. It may be so personalized to the subjectivity of each of us that we’ll have to re-imagine the concept of an “audience.” New mediums change the way we think. In the same way cinema has changed the way we imagine the world (for example, novelists now write “cinematically”), AI will change the way we understand our subjectivity.

When it comes to AI and art, we’re still only at the moment when all we can see is a flickering horse-drawn bus coming toward us. Much more playing, experimenting, critiquing, and, most importantly, failing, will be required to invent the language of AI as a medium for art.¹⁶ But one thing is certain: Just as the cinematograph transformed our relationship with actuality, the noematograph will transform our relationship with subjectivity.

1. From the program by the Lumière brothers. Available at
2. There are many ways to think about AI and art, and I want to be clear upfront on what I won’t be talking about. I won’t talk about the many ways that AI may be a threat — the environmental impact, the addiction to easy empathy, the damage to existing artistic communities and ecologies, the potential for debasement and degradation of human creativity, and so on. Not because these things aren’t important, but because there are many other thinkers better qualified and eager to talk about them already. Nor will I discuss the possibility of AI gaining consciousness — some variation of the Singularity — such that Homo sapiens must reckon with a truly new form of consciousness. That’s also very important, but if that were to occur, we would have much more interesting things to talk about than AI-generated art.
3. The invention of cinema is full of self-references like this, where a camera captures other cameras. The Lumière family business was photography, and their factory, featured in La Sortie de l’Usine Lumière à Lyon, manufactured photographic equipment.
4. Myrent, G. “When Movies Began and No One Came.” The New York Times, December 29, 1985. Méliès was not present at the public screening on December 28, 1895. Instead, he saw these films at a private showing for directors of entertainment venues around Paris the day before. See Cinémathèque Méliès: Lettre d’information, No. 37, June 2013, p. 7.
5. Sometimes, the best way to understand a new medium is to “translate” material from an existing medium into it.
6. Vaswani, Ashish, et al. “Attention Is All You Need.” Proceedings of the 31st International Conference on Neural Information Processing Systems, edited by Isabelle Guyon et al., vol. 30, Curran Associates, Inc., 2017, pp. 6000–6010. arXiv,
7. All the AI experiments were done with ChatGPT (4o) in January 2025. The output has been edited for length.
8. Well, still not exactly “correct.” By 古人, Chen Zi’ang likely meant more than “sages,” something closer to “wise sovereigns of the past who listened to talented individuals” (and he perhaps even had in mind one specific ancient ruler connected with the terrace he was ascending). Current AI models still mess up on these details.
9. I do have quibbles with all of them, but these are at the level of aesthetics, not the meanings, given the interpretive frameworks necessitated by the agendas behind them.
10. “Translating the Odyssey: How and Why,” Emily Wilson’s lecture at the University of Scranton, December 18, 2018. https://www.youtube.com/watch?v=MyLs9VoY9tc
    Now, suppose you keep on pushing ChatGPT to give you more interesting/unconventional translations. In that case, it will comply much like the hallucinatory image generators in those “make the bodybuilder more muscular” trends (because these models tend to please you sycophantically). At some point, the translations start to become — to my mind at least — impossible to justify. (The same charge, of course, has been levied against human translators in many cases.) I’m old-fashioned enough to believe that meaning exists, and you can’t just make any text mean anything you like. Just how we draw this line between an acceptable rendition of (some of) the truths in a text and an unacceptable injection of non-truths into a text is a question that, perhaps fortunately, no AI can help us resolve.
11. Sometimes, this capturing of subjectivity is even explicit, such as when the CIA constructs chatbots to stand in for world leaders and interrogates them. See
12. This is distinct from situations where AI is used as a tool — a brainstorming partner, a beta reader, a critic, a tutor, etc. — to help humans do what they already do, but better, at a higher quality, with more thought and care. That would be lovely, too, and may lead to more interesting books, pictures, and movies, but it would not be using AI as a medium in itself.
13. You can find countless YouTube videos for “How to use AI to write a story.” Try to find some (any) for “I really loved reading this AI-written story.”
14. There is another possible contributing cause to the boring output: the thick layers of “don’t-let-the-model-say-or-produce-anything-that-will-get-the-company-in-trouble” code that all these generative AI models are wrapped in. Unfortunately, whenever you try to make a model’s output more conventional (and thus more clichéd), you also make it less interesting. That’s kind of baked into the nature of meaning-making and subjectivity.
15. Indeed, sometimes the stances that the user wishes the AI to take on can be so transgressive (by the standards of the corporate/governmental overlords of these models) that “jail-breaking” is required.
16. Here’s something to watch out for: The development of the language of cinema was the result of experimentation by thousands of independent creators who were all free to push the medium in whatever direction they liked; however, AI may end up being controlled by a few powerful corporations (enabled by even fewer paranoid governments). Centralization and concentration of control are not good for the development of new mediums.