Case Study 1: The Portrait That Sold for $432,500

The Event

On October 25, 2018, Christie's auction house in New York sold a painting titled Portrait of Edmond de Belamy for $432,500 — more than 40 times its estimated price. The painting depicted a blurry, somewhat unsettling portrait of a man in dark clothing, rendered in a style that evoked old master paintings while remaining distinctly strange. The canvas was signed, but not by a human name. In the lower right corner, the signature read:

min G max D E[log(D(x))] + E[log(1-D(G(z)))]

It was a mathematical formula — the loss function of a Generative Adversarial Network.

The painting was created using a GAN trained on a dataset of approximately 15,000 portraits spanning the 14th through the 20th centuries. The GAN learned the statistical patterns of portraiture — faces, poses, backgrounds, color palettes, brushwork — and generated a new image that recombined these patterns in a configuration that had never existed before.

The sale made international headlines. It was covered by major newspapers, art magazines, and technology outlets. It was described as the first AI-generated artwork to be sold at a major auction house. And it triggered a fierce debate that, years later, remains unresolved.

The Creators — Or Were They?

The painting was brought to auction by a French art collective called Obvious, consisting of three members in their mid-twenties: Hugo Caselles-Dupré, Pierre Fautrel, and Gauthier Vernier. Obvious didn't build the GAN from scratch. They used code created by Robbie Barrat, a then-19-year-old AI researcher and artist who had published his GAN training code as open-source software on GitHub. Barrat had been training GANs on portrait datasets and sharing the results as art projects.

Obvious used Barrat's code (or a close adaptation of it), trained the model on their own curated dataset of historical portraits, generated images, selected the ones they found most interesting, printed one on canvas, framed it, and brought it to Christie's.

This chain of creation raised the first authorship question: Who made this painting?

The GAN generated the image through a computational process. But the GAN had no intention, no aesthetic preference, and no concept of what a "portrait" means. It produced an image that was statistically consistent with its training data.

Robbie Barrat wrote the code that made the generation possible. He didn't choose to create this specific image, wasn't involved in the auction, and didn't receive any of the proceeds. He publicly expressed frustration that Obvious received credit and payment for work that depended heavily on his open-source contribution.

Obvious curated the process — choosing the training data, selecting the output, deciding to present it as art, and bringing it to the market. They exercised judgment at every stage, but they didn't write the code or generate the specific pixels.

The 15,000 portrait artists spanning six centuries created the works that the GAN learned from. They were long dead (most of them), had no opportunity to consent, and received no credit in Christie's promotional materials.

Christie's made the framing choice to present this as a landmark event in art history, generating media attention that drove the price up. Their marketing played a significant role in the painting's cultural significance.

The Debates

Debate 1: Is This Art?

The art world's response was divided.

Yes, it's art. Some critics and collectors argued that the history of art is full of new techniques that were initially dismissed. Photography was called "not art" because a machine produced the image. Ready-mades (Duchamp's urinal) were called "not art" because the artist didn't make the object. Conceptual art was called "not art" because there was nothing to look at. Each time, the definition of art expanded. AI art, by this logic, is simply the next expansion.

The Portrait of Edmond de Belamy provoked a genuine aesthetic response in viewers — curiosity, unease, fascination. It raised questions about portraiture, identity, and the nature of artistic creation. If art is about provoking thought and feeling, the piece succeeded.

No, it's not art. Others argued that the painting lacked the essential ingredient of art: human intention. The GAN didn't intend to create anything. It didn't struggle with artistic choices. It didn't have something to communicate. The "interesting" qualities of the image — its blurriness, its uncanny quality — were artifacts of the GAN's limitations, not artistic choices.

By this view, presenting a GAN output as art is like presenting an interesting cloud formation as sculpture. It might be visually striking, but it wasn't created — it happened.

It's complicated. A third camp argued that the question "Is this art?" was the wrong question. The more interesting questions were: Whose art is it? What does it mean that a computational process can produce something that looks like art? And what does the art world's willingness to spend $432,500 on it reveal about the art market?

Debate 2: Whose Art Is It?

The authorship question was more contentious than the art question.

Obvious presented themselves as the artists. They argued that their creative contribution — selecting the training data, curating the outputs, making the decision to present the work as art — was the artistic act. They compared themselves to photographers (who use a camera to capture, not create) or to Andy Warhol (who used silk-screening techniques that were partly mechanical).

Robbie Barrat argued he was uncredited. The code that generated the image was largely his. He had been creating and sharing AI art for months before Obvious came along. He didn't claim to be the sole artist, but he felt that his contribution deserved recognition and that Obvious had benefited from his open-source work without adequate attribution.

The broader AI art community pointed out that thousands of researchers contributed to the development of GANs. Ian Goodfellow invented the architecture. Hundreds of researchers improved it. The open-source community made the tools accessible. The "authorship" of AI art is distributed across a vast network of contributors, most of whom are invisible.

Debate 3: What Does the Price Mean?

The $432,500 price tag was itself a subject of debate. Was the buyer paying for the aesthetic quality of the image? For the novelty of owning a "historic" AI artwork? For the concept — the idea that AI can create art? For the bragging rights?

Critics pointed out that the painting's aesthetic quality was, by traditional standards, unremarkable. The blurriness and distortion that made it "interesting" were computational artifacts, not artistic choices. Similar (or arguably better) AI-generated images could be produced in seconds by anyone with access to the same tools.

The price, these critics argued, reflected the art market's love of novelty and narrative more than the work's intrinsic quality. The story — "first AI artwork at a major auction house" — was what sold, not the painting itself.

Supporters countered that art has always been about more than technical skill. Marcel Duchamp's Fountain (a urinal signed with a pseudonym, submitted to an art exhibition in 1917) was neither technically impressive nor aesthetically beautiful. It was important because of the questions it raised about what art is. The Portrait of Edmond de Belamy played a similar role for a new era.

What Happened Next

The sale opened the floodgates. In the years following, AI-generated art became increasingly mainstream:

• Artists like Refik Anadol created large-scale installations using AI-generated visuals that were exhibited in major museums.
• AI art tools like Midjourney and DALL-E made image generation accessible to millions of non-artists.
• In September 2022, an AI-generated image won first place in the digital arts category at the Colorado State Fair, sparking another wave of debate when the human competitor, Jason Allen, described himself as the artist.
• Lawsuits over training data mounted as visual artists organized to challenge the use of their work in training datasets.

The Portrait of Edmond de Belamy didn't settle any of the questions it raised. But it put those questions on the public agenda in a way that technical papers and academic conferences hadn't.

Discussion Questions

1. Who deserves credit (and payment) for the Portrait of Edmond de Belamy? Make a case for one of the following: Obvious, Robbie Barrat, the GAN itself, the historical portrait artists in the training data, or "no one."

2. Obvious compared their role to that of a photographer — someone who uses a technological tool to create images. Is this a valid comparison? What are the similarities and differences between using a camera and using a GAN?

3. If AI can generate portrait-like images in seconds, what happens to the cultural value we assign to portraiture? Does the ease of generation diminish the significance of human-painted portraits, or does it make them more valuable by contrast?

4. Robbie Barrat published his code as open-source — freely available for anyone to use. Does that mean Obvious did nothing wrong by using it? What are the ethical expectations around open-source software, and how do they intersect with artistic credit?

5. Apply the "authorship gradient" concept from the chapter. Where would you place the Portrait of Edmond de Belamy on the gradient? Would your placement change if Obvious had written the GAN code themselves?

Connection to Your AI Audit Report

Consider whether the AI system you're auditing generates or could generate creative content. If so: - Where on the authorship gradient would the output fall? - Who would claim credit for the output in practice? Who should claim credit? - Were the creators of the training data acknowledged, compensated, or given the opportunity to opt out? - How does the commercial context (who profits) shape the authorship question?