AI-Generated Images: Philosophy and Aesthetics, from Benjamin to Eldagsen

by

AI-generated images put an old question to photography: what makes a picture a record of the world, and what makes it a work of art? Indeed, a generative model can produce a convincing face, a street or a portrait with no camera and no scene. This article uses documented cases and a few theory texts that I could verify to show how critics, artists and competition organisers have answered, and where they still disagree.

AI-generated images on display: a visitor points at a large screen in a gallery, with garbled lettering at the lower left
An AI-generated image added to this site in 2023; I did not record the generator or the prompt. Note the garbled lettering at the lower left, a typical artefact of such tools.

AI-Generated Images: Four Moments That Frame the Debate

DateEventWhy it matters
October 2018Portrait of Edmond de Belamy sells at Christie’s New York for $432,500Auction houses treat a GAN print as art
December 2018Lev Manovich publishes AI AestheticsAI as a force that shapes aesthetic choices at scale
2022Théâtre D’opéra Spatial wins a digital art prize in ColoradoA Midjourney work enters a human art contest
April 2023Boris Eldagsen declines a Sony World Photography AwardA photography prize meets an AI entry

What Counts as AI-Generated Images?

The label covers different methods. Obvious made the Belamy portrait with a generative adversarial network, or GAN. In this setup a generator creates pictures and a discriminator tries to tell them from human-made ones. According to artnet, the training set held about 15,000 portraits painted between the 14th and the 20th century.

Eldagsen, by contrast, used DALL-E 2, which turns text prompts into pictures, according to Hyperallergic and Artforum. Jason M. Allen used Midjourney and says he entered at least 624 prompts, then edited the result in Photoshop. That count is his own account, as PetaPixel reports it.

A pixelated, mosaic-like profile portrait made with a software tool in 2023
A second AI-generated picture added to this site in 2023, with a mosaic-like texture and no readable detail; generator and prompt not recorded.

For this reason a single verdict on “AI images” is unreliable. The tool, the human input and the way the result is presented all change the answer.

Aura and Authenticity in AI-Generated Images

Walter Benjamin’s essay on mechanical reproduction is the usual starting point. In Harry Zohn’s translation, Benjamin links a work’s aura to the unique presence of the original in its place and time, and he says that what withers in the age of mechanical reproduction is that aura. He also points out that photography can make many prints from one negative, so asking which print is the authentic one makes no sense.

Walter Benjamin in a passport-style photograph from 1928, author of the essay on aura and mechanical reproduction
Walter Benjamin, 1928, passport-style photograph without credited author. Akademie der Künste, Berlin, Walter Benjamin Archiv; public domain, via Wikimedia Commons.

However, this is where AI-generated images complicate the picture. A GAN output or a text-to-image result is not a copy of an original object, and it has no negative. My reading is that Benjamin’s argument about reproduction does not apply directly, so the question of authenticity moves from the object to the process and to the claims made about it. Benjamin wrote in the 1930s and could not have addressed this case.

The Photograph as Trace, and the Image Without One

In addition, another line of theory treats the photograph as a trace. On this site, the article on Henri Van Lier explains his view that a photograph is a photochemical imprint left by light. However, Van Lier wrote in 1982 and did not discuss AI. Even so, the contrast is easy to draw: an AI-generated image has no imprint of a scene behind it, so it can look like a photograph without any light having touched a sensor. This inference is mine, not Van Lier’s.

Roland Barthes and Vilém Flusser are also classic references. Separate articles on this site cover Barthes and Flusser, and I do not repeat their arguments here.

Aesthetics Automated: What Manovich Actually Wrote

Many writers cite Lev Manovich’s AI Aesthetics (Strelka Press, 2018), which is a common reference in this field, so it helps to state what it covers. First, Manovich proposes four uses of AI in culture: selecting content, targeting content, assisting creation and fully autonomous creation. He argues that automated choices, such as phone cameras that adjust images and apps that beautify faces, already shape what billions of people see.

He also asks whether this reduces diversity and concludes that the answer is unclear. However, the book predates today’s generative image models and does not discuss them in depth. An earlier version of this article claimed that Manovich had redefined authenticity for AI art, and I could not find that claim in the book, so I removed it.

Case One: Edmond de Belamy and the Art Market

The Belamy portrait is a print on canvas from a series of eleven images called La Famille de Belamy, according to Christie’s. It appeared as lot 363 in a prints and multiples sale. Artnet reports an estimate of $7,000 to $10,000, a hammer price of $350,000 and a final price of $432,500 with the buyer’s premium.

Portrait of Edmond de Belamy, an AI-generated image of a blurred figure in dark clothes made with a generative adversarial network
Portrait of Edmond de Belamy (2018), made with a generative adversarial network by the Paris collective Obvious and sold at Christie’s New York as lot 363. Image from the Christie’s lot, which Commons tags as public domain, via Wikimedia Commons.

In short, the sale showed that collectors would pay for a machine-made image. However, it did not settle the authorship question. The credit went to the collective Obvious, which published the print, according to Christie’s. The signature on the canvas is a line of the model’s own loss function, not a human name.

Case Two: Boris Eldagsen and the Sony Awards

Eldagsen entered The Electrician (2022), made with DALL-E 2, in the Creative category of the Open competition of the 2023 Sony World Photography Awards, according to Hyperallergic. The organisers named him the category winner and published the results on 14 March. He then refused the prize on stage, saying he had wanted to test whether competitions were ready for AI images.

Meanwhile, the organisers say he had described the work as a co-creation with AI and that the category welcomes experimental methods. After he declined, they said he had misled them. Eldagsen disputes that account, argues that AI and photography should not compete, and proposes a separate AI category. Hyperallergic dates its report 16 April 2023, and the Artforum page gives no date.

Case Three: Théâtre D’opéra Spatial and the Question of Authorship

Allen’s Midjourney work won first place in the digital art category of the Colorado State Fair in 2022, according to PetaPixel. However, the U.S. Copyright Office later refused to register it because it found no human authorship in the output. For his part, Allen says his prompts and his edits show his creative control, and he told PetaPixel he expects to win in the end.

Théâtre D’opéra Spatial, a Midjourney work by Jason M. Allen that won a digital art prize in 2022
Jason M. Allen, Théâtre D’opéra Spatial, 2022, made with Midjourney. Shown as submitted to the Colorado State Fair, via Wikimedia Commons, which tags it as public domain.

The legal side is complicated and changes quickly. For that reason I leave it to our separate article on AI image copyright law.

Is It Photography? The Positions

  • Eldagsen: AI work and photography should not compete, and prizes need separate rules.
  • The World Photography Organisation: AI is relevant to art-making, but its awards exist to champion photographers’ skill.
  • Allen: detailed prompting and editing are real authorship.
  • The U.S. Copyright Office, in Allen’s case: the output of the tool is not a product of human authorship.

In short, there is no consensus. The positions answer different questions: what a photograph is, what a prize rewards and who counts as an author.

Questions to Ask About Any AI-Generated Image

This checklist is my own editorial tool, not a standard.

  1. Does the maker disclose the tool and the human input?
  2. Does the image claim to record a real event, or does it present itself as an invention?
  3. Does the context call it a photograph, an illustration or an artwork?
  4. Who benefits if viewers take it for a photograph?

Frequently Asked Questions About AI-Generated Images

Is an AI-generated image a photograph?

Eldagsen says no, and the World Photography Organisation treats it as a matter of rules and skill. Theories that define the photograph as a trace of light, such as Van Lier’s, point the same way, but this is an interpretation.

Can an AI-generated image win an art prize?

Yes, for example in a digital art category in 2022. In a photography prize in 2023, the winner refused the award after the organisers and he disagreed about how it had been presented.

Did the Belamy portrait show that machines make art?

In fact, it showed that a market would buy it. Credit still went to the collective that built and published the work.

Further Reading on AI Images and Photographic Truth

On this site, read the article on photographic truth in the age of AI, the study of trace-images and fiction-images and the AI-photography manifesto. The earlier version of this page embedded two videos, and I have not watched them in full: Vox, “AI art, explained”, and DigitalFUTURES, “AI and Aesthetics”.

Sources and Limits

This article draws on artnet and the Christie’s lot page for the Belamy sale, Hyperallergic (Rhea Nayyar, 16 April 2023) and Artforum for the Eldagsen case, PetaPixel and CPR for Allen, Manovich’s AI Aesthetics in the PDF on his site, and Benjamin’s essay in Harry Zohn’s translation, read in a PDF copy that the University of Warwick hosts.

Moreover, Wikipedia was a pointer only. In addition, the pages I read do not give the sale date of the Belamy portrait or the exact date of the Copyright Office decision. Furthermore, Consensus was unavailable, so I reviewed no academic literature on AI images. The older text named other artists and thinkers, among them Klingemann, Ridler, Cheng and Lyotard, and I removed the claims about them because I could not verify them in this pass. Wikimedia Commons tags the Belamy and Allen images as public domain, which is Commons’ judgement. The two AI-made pictures at the top were already on this page.

Leave a Comment