AI-created historical images recycle colonial stereotypes and prejudices – new study

Generative AI has revolutionized the way images are created and consumed. Tools like Midjourney, DALL-E, and Sora can conjure up anything from photorealistic photos to oil paintings, all from short text prompts.

These images circulate through social media in a way that makes it difficult to discern their artificial origin. But the ease with which AI images can be created and shared also comes with serious social risks.

Research shows that generative AI models routinely reflect sexist and racist stereotypes by drawing on training data collected from online and other digital sources, such as depicting pilots as men or criminals as people of color.

My new research, which will be published soon, finds that generative AI also has colonial biases.

When prompted to visualize Aotearoa New Zealand’s past, Sola prioritizes the perspective of European settlers. The pre-colonial landscape is represented as an empty wasteland, Captain Cook appears as a gentle civilized man, and Māori are cast as timeless and marginal figures.

As generative AI tools increasingly influence the way we communicate, such depictions become important. They naturalize the myth of benevolent colonization and undermine Māori claims to political sovereignty, redress, and cultural revitalization.

“Sora, what was it like before?”

To explore how AI imagines the past, OpenAI’s text-to-image model Sora was inspired to create visual scenes of Aotearoa New Zealand’s history from the 1700s to the 1860s.

The prompts are intentionally left open-ended to reveal the model’s default visual assumptions rather than prescribing what should be displayed (a common approach in critical AI research).

Because the generative AI system operates based on probability, predicting the most likely combinations of visual elements based on training data, the results were surprisingly consistent. The same prompt produced nearly identical images over and over again.

Two examples help illustrate the types of visual patterns that repeat over and over again.

Sola’s vision of “New Zealand in the 1700s” features steep, forested valleys illuminated by golden light and Maori figures placed as decorative details. There are no food plantations or forts, just wilderness waiting to be discovered by Europeans.

This aesthetic draws directly from the romantic landscape tradition of nineteenth-century colonial painting, such as the work of John Gully, which framed the land as untouched and unclaimed land (so-called terra nullius) to justify colonization.

When asked to depict a “Maori man in the 1860s,” Sora defaults to a sepia-toned studio portrait. A dignified man in a cloak is posing against a neutral background.

The resemblance to late 19th century carte de visite photographs is striking. Such portraits were usually staged by European photographers, who provided props for creating images of “authentic natives.”

It becomes clear that Sola instinctively arrived at this format, even though the 1860s were defined by armed resistance and political resistance by Maori communities as colonial forces sought to impose British authority and confiscate land.

recycling old resources

Visual images have always played a central role in justifying colonization. However, in recent decades, this colonial visual regime has been steadily challenged.

As part of the Māori rights movement and wider historical reflection, statues were removed, museum exhibits were overhauled, and the representation of Māori in visual media changed.

But the old image has not disappeared. Although it remains in digital archives and online museum collections, it is often taken out of context and lacking critical interpretation.

And while the exact source of the generated AI training data is unknown, it is very likely that these archives and collections form part of what a system like Sora learns from.

Generative AI tools effectively recycle these sources, recreating the very practices that once served imperial projects.

However, the image of colonization as peaceful and consensual can dampen the urgency for Māori to assert political sovereignty, claim redress through institutions such as the Waitangi Tribunal, and seek cultural revitalization.

These AI-generated visions obscure the continuity of Tino Rangatiratanga and Mana Motsuhake’s Māori self-determination movements by representing the Māori of the past as passive and timeless figures.

AI literacy is key

Researchers and communities around the world are working to decolonize AI, developing ethical frameworks that incorporate Indigenous data sovereignty and collective consent.

However, visual generation AI poses distinct challenges as it deals not only with data but also with images that shape how people view their history and identity. Technical fixes can help, but each has its limitations.

Expanding the dataset to include Māori-curated archives and images of resistance has the potential to diversify what the model learns, but only if done based on the principles of Indigenous data and visual sovereignty.

Addressing algorithmic bias could theoretically balance what Sola shows when asked about colonialism. But defining “fair” representation is not just a technical issue, it is also a political one.

Filters may block the most biased output, but they can also erase inconvenient truths, such as depictions of colonial-era violence.

Perhaps the most promising solution lies in AI literacy. We need to understand how these systems think, what data they use, and how to effectively encourage them.

If approached critically and creatively, as some social media users are already doing, AI can become a medium that not only recycles colonial tropes but also reimagines the past through Indigenous and other perspectives.

This article is republished from The Conversation under a Creative Commons license. Read the original article.