AI Everywhere in the Building, Never on the Wall

In December 2022, one of the large stock-image platforms began accepting AI-generated pictures, labelled as such. Two researchers, Samuel Goldberg at Stanford and H. Tai Lam at UCLA, watched what happened next.¹ The supply of images jumped. Sales rose by more than a third. Buyers showed a clear preference for the AI pictures. And the number of human artists selling on the platform fell by almost a quarter. The researchers’ summary is hard to argue with: consumers won, artists lost.

I run a museum built almost entirely with AI,² and I think the study is right. I also think almost everyone is drawing the wrong lesson from that result, because one word goes unexamined: stock.

A stock image has a job. It is bought to communicate someone else’s topic: the blog post needs a header, the deck needs a visual, the campaign needs a mood. It is chosen by search term, judged only by fit, and discarded when the topic changes. Nobody frames it. The buyer was never looking for a picture at all; they were looking for a way to say something faster. For that job, a machine that produces exactly the image you describe, in seconds, for nearly nothing, is simply the better tool. The study’s buyers were not wrong, and the honest response is not to scold them.

I know this from my own hands. On my consulting site I paired my essays with real artworks I love, chosen for the resonance I saw between the painting and the argument. The feedback from business readers was consistent: they did not see the connection. They are not wrong either. If I generated an image of exactly the essay’s idea, a supply chain as a river system, a decision as a fork in a mountain path, it would communicate better, because communication is the job you hire an illustration to do, and a made-to-order image does it better than a four-hundred-year-old one. When I want an image to serve my topic, the machine serves best.

But serving your topic is only one of the two jobs a picture can do, and it is the lesser one; call it communication.

The other job is the opposite of service. A real work of art was not made for your brief. It was made by someone with their own stake in being alive: a woman who talked the Paris police into letting her sketch at a horse market,³ a brass caster working for a king whose palace would later be burned,⁴ a printmaker putting beauty on ten thousand ordinary walls.⁵ When you put a work like that next to your thinking, it does not illustrate your idea. It resists it. It arrives with someone else’s whole answer to some other question, and the friction between their answer and your question is where the provocation happens. Thought is not provoked by images of itself. It is provoked by contact with another mind. That is why art meant to stretch you is best built on someone else’s creativity, and why an image generated from your own prompt can almost never do this particular work: there is no other stake in it to meet. You asked for it. It agrees with you. Agreement never stretched anyone. Researchers who tried to use generated imagery to help people picture unbuilt futures hit the same wall: across 480 images, prompting for the future kept returning the past, the same dominant faces the training data already held.⁶

So the two jobs pull in opposite directions. Communication wants the image on-topic, and the machine is the master of on-topic; provocation wants the image off-topic in exactly the way another human life is off your topic. There, the machine has nothing to offer: there is nobody home.

This is why our museum is built the way it is, and the way it is built surprises people on both sides of the AI argument. We are AI-native to the bone. Our collection is served to AI assistants through an open protocol built on the Model Context Protocol,⁷ so that anyone’s AI can walk our galleries. Our audio tours are narrated by a synthetic voice, reading words we wrote. The engineering is AI-assisted end to end. By any fair measure, we are an AI company. And no AI-generated image will ever hang on any surface of ours.⁸ Not the site, not a post, not a poster, not a page of a book. Both facts follow from the same view of what the technology is for: the machine is a superb builder and a superb servant of topics, and it is not a maker of the thing we exist to show.

The economics do not frighten us either, and the study explains why they should not. What AI displaced on that platform was communication work done by human hands, and that work is going the way typesetting went; pretending otherwise helps no one who needs to pay rent. What AI cannot displace is the work with a person in it, but that survival is not automatic. It depends on whether the two jobs stay distinguishable, and the burden of keeping them distinguishable falls on institutions, not on viewers scrolling at speed. A panel we sat in during Art Basel this June arrived at the same point from the other direction: the real risk of AI here may not be fakery but volume.⁹ Flood the world with synthetic images, and telling the real from the generated gets expensive; the cost starts to look like a quiet kind of censorship. Millions of real human works are within reach, most of them free to see, many free to reproduce. For someone choosing what to live with rather than what to communicate, the generated image holds one advantage: convenience. It is easy, and it asks nothing of you. Banning the easy thing is not our answer. We want to beat it at ease: make the real work as effortless to find, to verify, to bring home, as the synthetic one, and do the curation and the storytelling so the easy choice is also the true one.

The artists who lost their livelihood on that platform deserve better than a eulogy, and better than a boycott run on their behalf. The woman at the horse market was not rescued; she got her permit and painted. What artists like her deserve is institutions that can tell the difference between an image that says something for you and a work that says something to you, and that put their money, their engineering, and their walls behind the second. Consumers won and artists lost in the market for saying things faster. The market for being changed by another mind is still open. We intend to keep it that way, with every machine we have.

Sources

1. Samuel Goldberg & H. Tai Lam, “Generative AI & Creative Goods: Market Expansion, Crowd-Out, and Copyright,” SSRN working paper 5152649, reported in “When AI-Generated Art Enters the Market, Consumers Win, Artists Lose,” Stanford Graduate School of Business Insights

2. Open Museum

3. Rosa Bonheur, The Horse Fair, 1852–55, The Metropolitan Museum of Art

4. Plaque with equestrian ọ́bà (king) and attendants, Ìgùn Ẹ́rọ̀nmwọ̀n brass-casting guild, Kingdom of Benin, ca. 1540–70, The Metropolitan Museum of Art

5. Utagawa Hiroshige, Driving Rain at Shōno, from Fifty-Three Stations of the Tōkaidō, 1833, Cleveland Museum of Art

6. Laura Aline Bechthold, “Rendering New Futures or Enshrining Pictures of the Past? The Double-Edged Sword of Using AI-Generated Imagery for Anticipation,” Journal of Futures Studies, vol. 30 no. 4, June 2026

7. Open Museum MCP, the collection’s open protocol for AI assistants

8. Open Museum, Code of Ethics

9. “Art Without Artists, Texts Without Authors, Goddamn Algorithms Everywhere,” a panel at Art Basel, June 2026, moderated by Trevor Paglen, with Aria Dean, Josh Kline and Hari Kunzru. Threads as we heard them in the room, not quotes.