When Christie’s closed Augmented Intelligence on March 5, 2025, it did more than put AI-assisted images under an estimate. The 34-lot sale tested whether a major auction house could turn a contested medium into a legible market category—and whether buyers would pay for the artist’s process as much as for the image.
The result was substantial but not uncomplicated. Christie’s reported a sale total of $728,784, saying the figure surpassed the presale estimate. Its results page records 34 lots, while Christie’s press release says the auction attracted hundreds of bids. Independent reporting by Artnet counted 28 lots sold. That is a useful distinction: a headline total can describe the money that changed hands without implying that every experiment found a buyer.
A market built around different kinds of machine involvement
The sale’s title was broad enough to hold several generations of computational practice. Harold Cohen’s Untitled (i23-3758) carried an estimate of $10,000 to $15,000 and realized $11,340. Charles Csuri’s Bspline Men, estimated at $55,000 to $65,000, realized $50,400. Alongside those historical figures were works by artists using contemporary machine-learning systems, including Refik Anadol, Holly Herndon and Mat Dryhurst, Claire Silver, Botto, and others.
That mixture matters. “AI art” is often used as if it names a single technique, but the Christie’s catalog grouped together rule-based computer art, neural-image experiments, NFTs, video, and physical works. The sale therefore functioned less like a verdict on one medium than like a compressed argument about what counts as authorship when software participates in making the work.
The strongest prices still carried a human story
The top lot was Refik Anadol’s Machine Hallucinations – ISS Dreams – A, which sold for $277,200 against a $150,000–$200,000 estimate, according to Christie’s. The auction house describes the work as drawing on more than 1.2 million images from the International Space Station and satellites. That description gives the work a recognizable structure: a named artist, a defined dataset, a research premise, and a visual result.
Holly Herndon and Mat Dryhurst’s Embedding Study 1 & 2 also exceeded its high estimate, selling for $94,500 against $70,000–$90,000. Christie’s says the work emerged from a text-to-image model trained on altered images of Herndon, and notes its connection to the 2024 Whitney Biennial. In both cases, the technology is not presented as an autonomous author. It is part of a chain of choices that collectors can name, discuss, and value.
What the results leave unresolved
Christie’s said 37 percent of registrants were new to the house and 48 percent of bidders were Millennials or Gen Z. Those figures are marketing claims from the auction house, not an independent measure of the market’s health, but they do point to the audience the sale was designed to reach. The catalog’s range of estimates—from low-thousands works to Anadol’s six-figure result—also suggests an attempt to make AI-related art legible at multiple price levels.
The unsold lots complicate any triumphalist reading. A sale total of nearly three-quarters of a million dollars confirms that collectors will spend meaningfully on AI-linked work. It does not establish that the category has settled its arguments about training data, originality, or the difference between an artist using a model and a company selling access to one. Christie’s Augmented Intelligence was most revealing as a market test: it rewarded works with clear artistic identities and histories, while leaving the larger question—what exactly is being bought when the machine is part of the medium—open.
