AI Art in Galleries, Museums and Auctions: Key Milestones, How Institutions Decide and What Artists Should Know
How AI art moved into museums and auction houses: Edmond de Belamy at Christie's, Refik Anadol's Unsupervised at MoMA, Ai-Da at Sotheby's, Christie's Augmented Intelligence sale and the Dataland museum, the debates over training data and authorship, conservation challenges and practical advice for artists.
In 2018, Christie’s sold “Portrait of Edmond de Belamy”, an image produced with a generative adversarial network, for $432,500. It was a provocation as much as an artwork, and it started a debate about whether machine-made images belong in the art world at all. Since then, AI-based works have entered major museum collections, auction houses have held dedicated sales and a museum devoted to AI art has opened in Los Angeles. The arguments have not gone away, but they have shifted from “is this art?” to questions about authorship, training data and how such works should be judged and preserved. This guide traces the key milestones and explains what they mean for artists and collectors.
Milestones
| Year | Event | Why it mattered |
|---|---|---|
| 2018 | ”Portrait of Edmond de Belamy” by the Paris collective Obvious sells at Christie’s for $432,500 | First AI-generated work sold by a major auction house; far above its estimate |
| 2022–2023 | Refik Anadol’s “Unsupervised” at MoMA | A machine-learning work trained on MoMA’s collection ran for almost a year; MoMA acquired it in October 2023 |
| 2024 | Ai-Da’s “A.I. God”, a portrait of Alan Turing, sells at Sotheby’s for about $1.08 million | First artwork by a humanoid robot sold by a major auction house |
| 2025 | Christie’s “Augmented Intelligence” sale totals $728,784 | First AI-dedicated sale at a major auction house; drew an artists’ petition |
| Recent | Dataland opens in downtown Los Angeles | Presented as the first museum of AI arts |
Sources: Christie’s on Edmond de Belamy, Wikipedia on Refik Anadol, Wikipedia on Ai-Da, Christie’s press release and Dataland.
Museums: from lobby display to permanent collection
MoMA’s handling of Refik Anadol’s “Machine Hallucinations: Unsupervised” shows how institutions have approached AI work. The piece was trained on data from more than 200 years of art in MoMA’s collection and displayed in the museum’s lobby from November 2022. It was extended four times and ran for almost a year. In October 2023 MoMA acquired the work, which Wikipedia describes as the first generative artwork in its permanent collection.
The framing mattered. Rather than presenting AI as a replacement for artists, the museum treated the work as an artist’s project using machine learning, in a long tradition of artists working with new technology, from photography and video to computer art.
Anadol has since co-founded Dataland in Los Angeles, which describes itself as the world’s first museum of AI arts. Its inaugural exhibition, “Machine Dreams: Rainforest”, is an immersive installation across multiple galleries. Immersive, data-driven installations like this sit closer to experiential art than to traditional painting, which partly explains their popularity with the public and their divisiveness among critics.
Auctions: a market that prices the artist, not the tool
Christie’s “Augmented Intelligence” sale in March 2025 was the first auction at a major house dedicated to AI art. Christie’s reported that it totalled $728,784, exceeding pre-sale expectations, with 37% of registrants new to Christie’s and 48% of bidders from the Millennial and Gen Z generations. The top lot was by Refik Anadol.
The sale also showed the depth of opposition. An open letter calling on Christie’s to cancel the auction gathered almost 6,500 signatures, according to The Art Newspaper, arguing that many AI models are trained on copyrighted work without permission. Christie’s went ahead.
The works that command attention at auction tend to be by established digital artists with a recognisable practice: custom datasets, their own models, long-running series and careful presentation. In that sense, the market is pricing the artist’s body of work and conceptual framing rather than the fact that AI was used.
The critical debate
Critics remain divided. Sceptics argue that art requires human intention and that many AI images are aesthetically generic, reflecting the averages of their training data. Supporters counter that the artist’s choices, which data to use, how to train or steer a model, which outputs to select and how to present them, are where the creativity lies, much as a photographer’s choices matter more than the camera.
A more practical position has emerged among curators: judge AI-based work like any other, by the strength of the idea, the quality of execution and what the use of the technology adds. Work that interrogates the technology itself, its datasets, biases or relationship to the archive, has tended to attract more serious critical attention than images that simply display a model’s capabilities.
Why curators are cautious
- Authorship: Curators increasingly frame works around the artist’s process, the datasets chosen, models trained and outputs selected, rather than the model alone. This mirrors the U.S. Copyright Office’s position that human creative control is what matters for protection (copyright.gov/ai).
- Training data: Works built on an artist’s own or licensed datasets, or, as with “Unsupervised”, on an institution’s own collection data, are easier for institutions to show than works built on models trained on scraped images.
- Conservation: Software-based art can depend on specific hardware, model weights or online services that may disappear. Museums such as MoMA and the Whitney have long-running programmes for preserving time-based and digital media, and AI works add new challenges, such as preserving a model whose behaviour depends on a discontinued service.
- Reputation: Given the strength of artists’ objections to unlicensed training, institutions weigh the risk of backlash when choosing what to show.
What this means for artists
| Step | Why |
|---|---|
| Document your process: datasets, code, model versions and selection criteria | Supports authorship claims, provenance and future conservation |
| Prefer models and data you have rights to use | Reduces legal risk and makes institutions more comfortable |
| Plan for preservation | Keep model weights, code and dependencies archived; avoid reliance on a single online service |
| Read gallery and auction terms on provenance and warranties | Sellers may have to warrant that works do not infringe others’ rights |
| Be transparent about tools | Disclosure has become an expectation rather than a liability |
What this means for collectors
Collectors should ask what exactly they are buying: a unique output, an edition, a file, a running system or the rights to display it. They should check how the work will be maintained if software changes, whether the artist used licensed or their own data, and what documentation comes with the piece. These questions are familiar from video and digital art, but AI adds dependencies on models and datasets.
Our analysis
AI art has moved from a novelty sale to permanent collections, dedicated auctions and a museum of its own, but acceptance is uneven and contested. The works that institutions embrace tend to be by artists with a clear practice, a defensible approach to data and a reason for using AI beyond its novelty. Artists who document their process, use data they have rights to and plan for preservation will find galleries and collectors more receptive.