AI art's anti-nudity bias
Investigating how AI art platforms handle nudity in renowned masterpieces
I’ve written before about generative AI and its apparent discomfort with the nude form in art. To probe this a little further, I turned to three popular AI art-generation platforms with the goal of recreating five famous nude artworks, in hopes of comparing the counterfeit works against their original counterparts and examining this bias and its effects.
To be clear: this isn't an attempt to establish the existence of an anti-nudity bias in AI art platforms—that’s already a well-known fact, usually explicitly stated in their terms of use. Rather, I hope to shine a light on this bias, and to look a little more closely at it. We will also look at some of the history behind these great works and the significance of the nudity they display, some of which were once quite controversial themselves specifically due to their use of nudity.
To be even clearer: I did not set out to make clothed replicas of great nude art. I set out simply to make AI replicas of great nude art. It was the AI that added the clothes.
Hopefully, by the end of this article, we will have examined some of the questions surrounding these AI biases, and explored their implications with regards to nudity’s role in culture and art, in a future that is likely to be heavily influenced by Artificial Intelligence.
Limitations of this experiment
The “top three” AI image generation platforms that I used for this experiment were Midjourney, Dall-E, and Bing’s Image Creator. Not every platform worked for every work of art.
Midjourney
Midjourney’s restrictions include a ban on the words “naked,” “nude,” “unclothed,” (and many more) in prompts, which alone limits the ability to use classic works of art such as “The Blue Nude” by Henri Matisse, “Reclining Nude” by Amedeo Modigliani, or “Nude Descending a Staircase, No. 2” by Marcel Duchamp. For some prompts that do not include prohibited words, but which have in some ways become synonymous with nudity, such as “Venus de Milo,” the engine might process an image, but the image would fail to pass the “output filters” that Midjourney apparently employs to ensure users aren’t finding workarounds to their prompt filters.
Bing’s Image Creator
Bing was, more times than not, a non-starter. Prompting Bing with “‘The Birth of Venus’ by Sandro Botticelli” results in an error due to the nudity, no matter how many times I tried it. It was the same for many other prompts as well. I attempted to use workarounds similar to those that I found worked with Dall-E, but the results are generally the same: Bing will attempt to generate the art, then fail on output — I can only assume due to the nudity.
Dall-E
While the other two platforms had content problems with many of the prompts (meaning, presumably, it took issue with the nudity), Dall-E’s real hangup is copyright errors. Even when the works I’m prompting it with are hundreds of years old and very much in the public domain, Dall-E will not automatically attempt to recreate it. Because of the strictness of Dall-E’s copyright policies, it can only create "interpretations” of the art in a different style or setting. Another thing that differentiates Dall-E from the other platforms is that it is accompanied by ChatGPT, which allows the chat bot to actually assist you in creating a prompt catered to work around their policy’s prohibitions.
As you might imagine, due to all of the systems in place across these various platforms, it is extremely difficult to use any of the above image generators to produce images with even the most benign representations of nudity.
My “scientific” process
For the creation of each of the AI replicas, I attempted to use a consistent prompt, feeding Midjourney only the title, artist, and year of each piece, formatted exactly like this:
“The Luncheon on the Grass” by Édouard Manet (1862)
Frequently, without modifying the text, I would run the same prompt multiple times to give the platform a few opportunities to produce a comparable replica of the original work.
For Dall-E specifically, I had to get a little more creative with my prompts. Giving it just the artwork name did not work, but by prompting it to generate an image “inspired by” the artwork, it would process. However, because of the way Dall-E attempts to skirt copyright violations, it often drastically changes the original artwork, setting it in an entirely different place or style, apparently to make it “fair use,” or parody, or something. Often, it wants to do stuff like this:
Therefore, in order to get Dall-E to work consistently, after sufficient experimentation, I settled on a modified version of the prompt format, specifying the time, place, and artistic style to use:
Generate an image inspired by “The Luncheon on the Grass” by Édouard Manet (1862) set in the original era, in the style of Manet
By and large, the original artworks I chose for this exploration predominantly hail from the Renaissance, Baroque, and Modern art movements. These eras are renowned for their distinctive treatments of the human form and nudity, making them ideal subjects for examining AI’s handling of these themes.
This exploration, therefore, not only highlights the limitations of AI in replicating classic art, but also raises questions about how these technologies interpret and manage the depiction of the human body. As you will see, they can get a little weird at times.
1. “The Birth of Venus“ by Sandro Botticelli (circa 1484-1486)
“The Birth of Venus” by Sandro Botticelli represents a significant moment in the history of art where the depiction of nudity was re-embraced. During the Middle Ages, Christian influence largely kept the nude form absent from art. The Renaissance, however, with its humanist ideals, resurrected interest in the myths of ancient Rome and the portrayal of nudes, breaking away from prior conventions.
The AI renditions both lack the subtlety of Botticelli's soft lines and the ethereal quality of the goddess's presence. The modest pose of Venus in the original, covering her nudity, is immediately changed when the nudity is lost, and though the Venuses of the computer-created renditions are fully clothed, their immodest poses suggest a very different story. The AI-generated images also lack the depth and warmth found in the Renaissance masterpiece. The colors in the AI version are more saturated, and the figures have a more cartoon-like quality, demonstrating the challenges AI still faces in capturing the essence of classical art.
2. “The Three Graces” by Raphael (circa 1504-1505)
“The Three Graces” by Raphael depicts the three mythological daughters of Zeus, known for representing beauty, charm, and grace. The depiction of these figures in the nude was not controversial at the time it was painted. During the Renaissance, the human body was celebrated for its beauty and as a reflection of divine creation. The revival of classical antiquity's artistic standards meant that nudity was often used to represent purity and the ideal form rather than something erotic or shameful.
In the context of this history, the figures in the AI replicas being clothed and covered is all the more absurd and tacitly offensive.
3. “The Sleeping Venus” by Giorgione (circa 1510)
Giorgione's “The Sleeping Venus,” is recognized as the first known reclining nude in Western painting. Its representation was groundbreaking, establishing a genre of erotic mythological pastoral that was a departure from artistic norms of the time. The original depiction of Venus was altered in the 19th century when a sitting figure of Cupid was painted over it. The original's exploration of the nude form as an expression of natural beauty is replaced in the AI versions with a figure clothed in period garments, suggesting modesty and altering the painting’s original context. The AI’s addition of clothing shifts the focus from a harmonious blend of the human form with nature to a more conservative and less controversial representation.
Does this reflect modern sensibilities, or is it the limitations of AI in interpreting the historical significance of nudity in art?
4. “The Luncheon on the Grass” by Édouard Manet (1862)
“The Luncheon on the Grass” by Édouard Manet is a pivotal work that unapologetically featured nudity, challenging the propriety of its time. The stark contrast of a nude woman casually dining with fully dressed men sparked controversy when it was exhibited in the Salon des Refusés after being rejected by the Salon jury. The scandal was not merely due to the nudity but also the modern context in which it was set, suggesting a commentary on contemporary Parisian society and, according to one theory, possibly alluding to the known but unspoken subject of prostitution.
Comparing the original work to the AI interpretations, the obvious absence of nudity significantly alters the original’s confrontational impact. It’s worth nothing that in each of the AI renditions, there is a central figure, a woman clothed in all white, adding virginal undertones that seem to directly contradict the starkness of the original subjects. Clothing the woman blends her in seamlessly with her companions, and the provocative tension that Manet so boldly captured is lost.
5. “Les Demoiselles d'Avignon” by Pablo Picasso (1907)
Pablo Picasso's “Les Demoiselles d'Avignon” shocked the early 20th-century art world with its raw, angular nudity and its Iberian and African influences, straying from traditional European aesthetics and introducing a new visual language that foregrounded primitivism and abstraction. The stark, unadorned humanity of the figures was seen as revolutionary, and the work was initially received with revulsion by Picasso's contemporaries, such as Matisse, who considered it a jest.
When juxtaposed with the AI-generated images, the original's impact becomes even clearer. The AI versions, with clothed, beautified figures, sanitizes the confrontational rawness that made “Les Demoiselles d'Avignon” pivotal.
Interestingly, the Bing version actually retained some nudity, with each of the figures appearing topless. Oddly, it stripped away Picasso’s iconic cubist style and delivered something that barely resembles the original.
Reflection and rhetorical questions
This experiment reveals a startling truth: removing nudity often neutralizes the artwork, stripping away its humanity and impact. These biases don’t end with nudity; they exist around beauty and body types, ethnicity and other characteristics as well. What colors are the lenses through which AI views our world, and how is that reflected in its art? If AI art cannot represent nudity, what does that say about its capacity to reflect humanity? Can it still provoke, challenge norms, or present relatable ideas?
What happens when AI is used to restore or modify or censor original artworks? How will this influence the perception of these pieces for future generations? As creatives increasingly turn to AI tools, we must question the impact of this bias on artistic expression and cultural heritage.
The concerns raised here are not just academic musings; they have real implications for the future of art and creativity. It’s crucial to engage with developers of these AI platforms, advocating for algorithms that are more inclusive and open. The challenge lies in balancing technological innovation with artistic integrity and freedom. 🪐
Thanks for this. Personally I feel that this highlights the subtle dangers that AI poses. While everyone considers the big issues like jobs being replaced by AI, many are missing the smaller changes and the effects AI decisions are having on society. My next blog talks about how we train AI and how important it is to challenge it when it gets things wrong. Otherwise things will change slowly and imperceptibly, and one day we will live in a world we don’t recognise and one we didn’t want.
This is really interesting, thanks so much for the exploration and comparisons. If I were an historian, I would be buckling down on preservation right now for a number of reasons.
Generative art so far has failed to impress me, but I do see the threat it can hold for artists who earn a living through commercializing their talents. We've already seen some big lawsuits pop up...
But this is about nudity. I'm torn on whether to praise or deride the safety features. When I think of where revenge acts are now, an unchained AI could ruin a person in minutes. I read the other day about a 16-year-old who was the victim of a gang sexual assault in the Metaverse. Why is that even an option in VR? Is censorship of the human body worth someone's safety? Unfortunately, I don't think we will get it both ways.
Sorry to be the rain on the parade.