The South Sea Maiden and the Black Nazi

These seemingly unrelated phenomena share a common root: they reflect not technological problems, but social ones. Or social realities to be exact. The issues with historical representations of Pacific Islanders and today's AI outputs stem from the same source: The state of our civilization and the forces that govern it.

When early European explorers first encountered Tahiti in the late 18th century, they quickly developed narratives about the island's women that would live on for centuries. Captain Samuel Wallis, who reached Tahiti in June 1767, remarked: "The women are all handsome, and some of them extremely beautiful". His shipmate George Robertson went further, claiming some sailors "swore they never saw hansomer made women in their lives".

Louis Antoine de Bougainville, arriving ten months later, described an atmosphere of "unbridled sensuality" and called Tahiti "La Nouvelle Cythère" after the Greek island of the love goddess Aphrodite. His surgeon and naturalist Philibert Commerson published observations in the Mercure de France, provocatively describing Tahitian women as "sisters of the Graces, and entirely without clothing".

These early accounts shaped how the West would view the South Pacific for generations to come. The "South Sea maiden" became a powerful symbol of feminine sexuality, of a presumed state of nature, and of Western fantasies about liberation from societal constraints. This narrative persisted despite the fact that initial European contact with Tahiti involved violence, with locals killed by powder. Only after Tahitians admitted defeat did chiefs send women to pacify the British sailors.

What's often overlooked in romanticized accounts is that these early interactions were far from the hedonistic stereotype. The sexual transactions between European sailors and Tahitian women were primarily economic exchanges: Women offering themselves for trinkets, a commodity previously unknown to them and which quickly grew in demand. Male relatives and local chiefs were often the main beneficiaries of these transactions, not the women themselves.

Moreover, many of the females engaged in these encounters would have just entered puberty. Henry Ibbot of the HMS Dolphin noted that the women taken to the ships were very young and small, and German scientist Georg Forster similarly remarked upon girls who did not appear to be more than nine or ten years old. These uncomfortable realities were rationalized by claims that women in warmer climates matured earlier than Europeans.

But rationality can fly in the face of history, such as that where Google's Gemini Chatbot caused controversy by generating historically inaccurate images including Black Nazi soldiers, female Asian popes, and racially diverse Vikings. These images resulted from Google's attempt to correct for the lack of diverse representation in AI training data.

The images were mind-boggling not because they were technically flawed, but because they reflected an overcorrection for biases in the underlying training data. Like the historical written accounts of Tahiti, AI systems learn from the data they're exposed to: In this case, billions of images and texts created throughout centuries, which themselves reflect our history, power structures, and society with all its biases and inequalities.

When we ask AI models to create images of people in various professions without specifying race or gender, they typically default to stereotypes: scientists, musicians, pilots, architects, and chefs are almost invariably depicted as men, predominantly white men. This is due to a process called, in NLP jargon, "Generalization".

This bias isn't a technological problem but a societal one. The training data for AI programs reflects the historical dominance of white men and ongoing present-day inequalities. When Google attempted to correct these biases by ensuring more diverse representations, the result was sometimes incongruous or offensive images that seemed to distort historical reality.

Now, some elements explain the historical misrepresentations of South Pacific women and the contemporary issues with AI image generation: Our history, society's structure, our contradictory aspirations, and the aggregating process of AI.

First, training data for AI programs, like historical accounts of the South Pacific, reflects the historic dominance of white men. When James Cook visited Hawaiʻi, his surgeon's mate David Samwell found the young women "exceedingly beautiful" and compared them to "Venus rising from the waves". With so many lascive women, "there was hardly one of Us that may not vie with the grand Turk himself". These accounts, like AI training data, overrepresented white male perspectives and underrepresented others.

In fact, AI systems expose ongoing structural inequalities. Just as colonial explorers saw Pacific Island women through the lens of their own societal power dynamics, AI reflects contemporary disparities. For example, when prompted to create images of doctors, AI consistently generates images of white men because that's what predominates in the training data.

Also, balancing historical accuracy with aspirational ideals proves challenging for AI just as it did for journal keepers such as James Morrison, a boatswain's mate aboard the HMS Bounty. When Generative AI models are prompted to depict women, those representations typically reflect male priorities: Women are usually highly sexualized, often infantilized, or entirely subdued. Similarly, historical accounts of South Sea women frequently emphasized their sexual availability while downplaying their agency.

AI's aggregating process mirrors how historical stereotypes formed. By distilling patterns from billions of images, AI models develop generic representations that gravitate toward the mean of their training data. In the same way, early European accounts of Tahiti were generalized and simplified and created enduring stereotypes that persisted regardless of their accuracy.

Now, should AI models reflect reality as it is, with all its biases, or should they present an aspirational view of how society could be tomorrow? Some want the world depicted as more diverse than it actually is, while others prioritize historical accuracy. The tension between these goals produces outcomes that can seem troublesome or offensive.

Google's effort to project an idealized version of society met resistance because it was perceived as clumsy and because its aspirations weren't universally shared. The resistance to "AI wokeness" may reflect resistance to social and demographic changes or to perceived social engineering.

What makes these issues particularly challenging is that there's no simple technical fix. Content moderation is inherently difficult because there are no simple rules that achieve the nuance, contextualization, and balancing required for decisions about what to allow and limit. Attempts to correct biases through simple rules inevitably produce new problems, as seen in the Black Nazi images.

In fact, when we crafted prompts to generate the illustration of this article, we tested an image-generation model to see whether it could draw the two women with pilosity under their armpits and on their legs. While the system was able to draw bare breasts with some astute privacy, the rendered images never included the reality of a woman's pilosity. It was unlikely that censorship, rather than insufficient training data on hirsute South Sea maidens, was the cause. Tampering with model algorithms and weights may backlash in ways we would never expect. What if we fiddled with the AI model to teach it that a south sea maiden did have unshaved hairs? Today's social forces of society would have no problem acknowledging that.

So, the solution isn't necessarily more sophisticated AI, but a better understanding of the historical and social forces that shape our perceptions. Just as we need a nuanced understanding of how European accounts of South Pacific women reflected European preoccupations rather than objective reality, we need to recognize that AI systems will always reflect human values and biases.

When AI produces unexpected or weird results, it's holding up a mirror to our collective history and current social dynamics. The Black Nazi Problem gives us a window into how AI exposes our ongoing struggle with the ramifications of past inequality and the difficulty of balancing inherently conflicting goals like diversity and historical accuracy.

As we continue to develop and use AI systems, we should expect them to sometimes produce outputs that make us uncomfortable: Not because they're wrong, but because they reflect unaccustomed truths about ourselves and the societies we've built, and the ones we want to build. The question isn't how to make AI perfectly unbiased, but how to use it thoughtfully with awareness of the human values and historical contexts it inevitably reflects.