Forum

Christie’s AI Sale and the Rise of Training-Data Provenance

A ceiling of carved orange-brown square coffers with an ornate chandelier hanging at the left

On Thursday, Christie’s opens Augmented Intelligence, a sale built around art made with artificial intelligence, from early pioneers such as Harold Cohen to contemporary artists including Refik Anadol. It runs until 5 March. Before a single lot has sold, it has already produced one of the sharpest public arguments about AI art so far.

On 8 February an open letter appeared online calling on the auction house to cancel the sale. Addressed to the specialists running it, the letter has drawn thousands of signatures, including those of illustrators Karla Ortiz and Kelly McKernan, who are suing several AI companies. Its central claim is direct: “Many of the artworks you plan to auction were created using AI models that are known to be trained on copyrighted work without a license.”

Two kinds of AI art

Christie’s response is revealing. A spokesperson said the artists in the sale have established, multidisciplinary practices, and that they use AI “in a controlled manner, with data trained on the artists’ own inputs.” Some of the artists have said publicly that they train their own models from scratch on their own material.

Whether that description holds for every lot is something collectors and critics will test. But notice what the defense is built on. Christie’s did not argue that training on other people’s work is acceptable. It argued that these artists did not need to. The question of what a model learned from has become part of how the work is presented and sold.

That is a significant shift. A year ago, the conversation about AI art was mostly about outputs: does the image look original, is it any good, who pressed the button? Now buyers, critics and fellow artists are asking about inputs.

A court looks at inputs too

The legal system is moving in the same direction. On 11 February, Judge Stephanos Bibas ruled in a long-running case between Thomson Reuters and Ross Intelligence, a legal research start-up. Ross had used headnotes from Westlaw, Thomson Reuters’ legal database, to build an AI search tool. The judge found the headnotes protected by copyright and rejected Ross’s fair use defense, pointing to the commercial purpose of the use and the harm to the market for the original.

The decision has limits that matter. Ross’s tool was not a generative system, and the judge said as much; the ruling does not decide how courts will treat image or text generators. Still, it is the first significant US decision to reject fair use for copying content to train AI, and it reinforces the point the Christie’s letter makes in plain language: where training material came from is not a footnote.

Provenance as a creative asset

For artists and designers, this suggests a practical strategy. Treat the provenance of your training data the way a photographer treats model releases or a printmaker treats edition records: as part of the work’s integrity.

Artists who build their own datasets from their own drawings, photographs or archives have a story that is both ethically clean and genuinely interesting. Those who rely on large commercial models can still be precise about what they did: which model, which version, which of their own materials went in, and what the human contribution was at each stage.

What will not hold up much longer is silence. A collector who pays for a work, a client who commissions a campaign and a jury that awards a prize all have reasons to ask what sits underneath.

Questions juries are starting to ask

Competitions are already moving this way. The 3rd International AI Award 2025 asks entrants which models and versions they used, how they prompted them and what ethical considerations shaped the work, and it requires entrants to own or have permission for what they submit. Its Early Bird deadline falls on 15 March, ten days after the Christie’s sale closes.

Good answers to those questions are not legal boilerplate. They describe a practice. A sentence such as “I fine-tuned a model on 400 of my own ink drawings, then edited the results by hand” tells a jury something about authorship that no finished image can show on its own.

What to watch

The sale’s results will be read as a referendum on the market’s appetite for AI art, but the more durable signal may be in how lots are described. If catalog notes start to specify datasets as carefully as they list materials and dimensions, provenance will have become part of the medium. Artists who prepare for that now will be ahead of the conversation, not chasing it.