AI Art Goes to Auction: The Christie's Sale and Why Licensing Matters More Than Ever

A historic sale, a heated debate

In late February 2025, Christie's opened the doors on "Augmented Intelligence" — the first major auction at a top-tier house dedicated entirely to art made with artificial intelligence. The sale featured 34 works, including a large canvas being painted in real time by a robot built by artist and roboticist Alexander Reben, where each $100 bid prompted the machine to add another layer of oil paint.

The auction pulled in hundreds of thousands of dollars in bids. It also generated something else: an open letter signed by more than 6,400 artists demanding the sale be cancelled.

The objection wasn't to AI as a medium. It was to the data underneath it.

The core complaint

Reid Southen, a Michigan-based film concept artist who helped organize the letter, estimated that roughly a third of the auction's works were produced with generative models — he named Midjourney, OpenAI's Sora, Runway AI, and Stable Diffusion — that were built on datasets containing copyrighted images without licenses or compensation. Southen also reported that he and many peers had seen their incomes roughly halved over the previous two years as AI tools moved into commercial pipelines.

The argument was less about new tools and more about consent. Prestige institutions like Christie's confer legitimacy, and legitimizing models trained on unlicensed work, the signatories argued, makes the broader practice harder to challenge.

Christie's, for its part, framed the auction as a continuation of art history rather than a break from it. Nicole Sales Giles, the auction house's director of digital art, welcomed the debate and compared AI's influence to the long tradition of artists drawing inspiration from one another. Critics counter that human influence and industrial-scale ingestion of millions of images are not the same thing.

Two camps inside the same auction

What the controversy sometimes obscured is that AI artists do not all work the same way, and the auction itself made that visible.

On one side were artists building on their own material. Daniel Ambrosi, for example, fed his own photography of Central Park into Google's DeepDream — a model originally created to study how neural networks recognize images, not to generate art. He described his process as that of a bandleader: writing the composition, then letting the model improvise within it. Several other contributors trained or guided models using their own photographs, collages, and writing.

On the other side were works produced with commercial generative models whose training data is broad, opaque, or admittedly drawn from the open web. The distinction matters — even an artist who feeds only their own inputs into a tool cannot always know what the underlying model was trained on. The model itself carries a history.

The legal backdrop

The auction unfolded against a shifting legal landscape. On February 12, 2025, Thomson Reuters won a copyright case against a legal research firm that had used its materials to train an AI system without permission — an early but consequential ruling on whether unlicensed training qualifies as fair use.

Meanwhile, OpenAI had already acknowledged in a UK filing the previous year that it would be "impossible" to train leading AI models without copyrighted material — a striking admission that today's generative AI industry rests on a foundation of unlicensed data, and that the question of how to compensate the underlying creators has been deferred rather than answered.

Why this matters for Creative Commons

The Christie's controversy is, at its core, a licensing controversy — and that puts Creative Commons squarely in the conversation.

CC licenses were built to do something copyright law alone does not: give creators a clear, machine-readable, internationally recognized way to declare in advance what others may do with their work. They are not a workaround for copyright; they sit on top of it. A CC license is a copyright holder's standing grant of specific permissions, on standardized terms anyone can read.

In an era of generative AI, that clarity is structural rather than decorative. Models trained on indiscriminate web scrapes face genuine ambiguity about what permissions, if any, attached to the underlying material. Models trained on CC-licensed corpora do not have that problem — provided the license terms are actually honored.

A short map of Creative Commons licenses

For artists publishing digital work, the six standard CC licenses run from most permissive to most restrictive:

• CC BY — Use, adapt, and build on the work, including commercially, with credit.

• CC BY-SA (ShareAlike) — Same as CC BY, but adaptations must carry the same license. This is the license behind most of Wikipedia.

• CC BY-ND (NoDerivatives) — Share the work, even commercially, but no modified versions may be published.

• CC BY-NC — Use and adaptation only for non-commercial purposes, with credit.

• CC BY-NC-SA — Non-commercial use and adaptation with credit; derivatives must use the same license.

• CC BY-NC-ND — The most restrictive: share with credit, no commercial use, no derivatives.

Alongside these sits CC0, a public-domain dedication that waives copyright entirely where law allows.

What digital artists can take from the moment

For artists thinking about both visibility and protection, a few practical points fall out of the Christie's debate.

Be explicit about license terms. Posting work with no license signal does not place it in the public domain. It leaves your rights ambiguous, which is the worst of both worlds. Whether you choose all-rights-reserved, a CC license, or CC0, make the choice visible on the work and in its metadata.

Think about what your license actually permits when it comes to AI training. Standard CC licenses do not carve out AI training as a separate use, but the underlying terms still apply. CC BY-NC would not permit incorporation into a commercial training set. CC BY-ND would arguably not permit the kind of transformation that training implies. Several efforts inside and adjacent to the CC ecosystem are now exploring explicit signals — opt-outs, machine-readable preferences — aimed specifically at AI training.

Document your inputs. If you train or fine-tune your own models, keep records of the data you used and the rights you had to it. Provenance documentation is becoming a baseline expectation in serious AI art practice, and may carry legal weight as well as ethical weight.

Read the model's terms, not just the prompt box. As the open letter pointed out, an artist using only their own inputs can still be working through a tool trained on material they would never have licensed. Reading model documentation — and where possible, choosing models with disclosed training data — is now part of the craft.

A closing thought

Whatever one makes of the Christie's auction, the more useful question it surfaces is not whether AI art should exist — that argument is over — but on what terms. Licensing is the language in which those terms get written. Creative Commons has spent more than two decades building that language for human creators sharing with other humans. Adapting it for an era when machines are also reading the catalog is the next chapter, and it is being written right now, partly in courtrooms, partly in open letters, and partly in the choices individual artists make every time they upload a file.

For a catalog of openly licensed digital art, the takeaway is straightforward: the more clearly creators state their terms, and the more carefully institutions honor them, the less of this debate ends up needing an auction house to resolve it.

Source: Natalie Kainz, "Christie's AI art auction draws big-money bids — and thousands of protest signatures," NBC News, February 26, 2025.