AI art replicates inequity at scale. We need to learn about its biases – and outsmart the algorithm

In 1840, the French painter Paul Delaroche saw the first daguerreotype – an early photographic process – and proclaimed that “from today, painting is dead”. One can understand him feeling threatened by the technology, but we know now he was dead wrong: instead, that year marked the birth of the art of photography.

Throughout history, there’s been fierce debate about the demarcation between science and the arts. It continues in the form of the great AI debate. In February this year, the winning photo in an Australian photography competition was created entirely by artificial intelligence; months earlier, a man used text-to-image software to take out the top prize in the Colorado Art Fair; and in 2018, a rudimentary painting produced by AI sold through Christie’s for US$432,500.

Late last year, in the remote Huon Valley in Tasmania, a freelance visual designer was inspired to experiment. Meng Koach had been commissioned by a publishing house to create a cover image for my new book about bias in artificial intelligence.

Through the eyes of a 55-year-old journalist like myself, this is a brave new world. A contentious and controversial world. There’s a long history of legislation and regulation trailing advances in technology.

• Sign up for our rundown of must-reads, pop culture and tips for the weekend, every Saturday morning

The process is exciting and terrifying. Koach typed words into Midjourney AI, which runs on Discord – a gaming platform I’m familiar with from several screaming matches with one of my teens. “Ghosts in the machine.” “A cybernetic forest.” “The bias of the past being built into the future.” These fragments make scant sense out of context. But they’re enough to construct a book cover within seconds.

The software uses billions of online images, identified by digital labelling from innovations like alt text. But the question remains: do AI art generators copy or steal other artists’ work?

“Unlike copying/stealing, there’s also taking inspiration,” Koach contends. “This subconscious accumulation of seeing and remembering art from other artists will, at one point, become apparent and trigger inspiration. Isn’t this the machine’s way of ‘taking inspiration’?”

The artistic community seems split. Some celebrate the removal of barriers into this rarefied realm; anyone can be an artist these days. Others are joining class actions, accusing the tech companies of appropriating art without credit, consent or compensation.

As a journalist, I’m fascinated but horrified by these developments. My book rails against the prospect of automation displacing millions of jobs. Still, this is a crucial conversation to be having right now. Technology continues to trample the media landscape. Even respected magazines such as The Economist are using AI to create artwork.

But using these programs is not as simple as it seems. While Koach was “shocked” at how quickly the initial cover images appeared, they didn’t quite convey the themes of the book. And as we added in new prompts – “artificial intelligence designed by man”; “back to the 1950s”; “strong women working together” – a different problem emerged. Female robots were sexualised with large breasts and tiny waists. The algorithm misinterpreted “strong women” as “massive biceps”.

My book, Man-Made, is about stereotypical images and words in datasets being used to train algorithms. These baby biases become troublesome teenagers through machine learning. The bots are increasingly bigoted, like white supremacists neck-deep in conspiracy theory websites.

Apps like Stable Diffusion, Dall-E and Midjourney can imagine a deserted island in the style of Monet. But if you ask for images of a CEO, it’s generally an older white male. Nurses? Almost all female. And if you don’t specify skin colour, the bots default to white people.

While humans are inherently biased, technology is replicating inequity at scale. We need to master this technology before it enslaves us.

skip past newsletter promotion

after newsletter promotion

Koach is keen to reframe AI art as simply another design tool. “The artist and the technology are intertwined throughout the creative process,” he says, and new roles are popping up in the sector too. Prompt Driver, anyone …?

Technology tends to move in one direction: its owners amass obscene profits while ethics is a mere afterthought. So what of the copyright issues? According to the Arts Law Centre of Australia, it’s still unclear. There are calls for AI models to be trained only on images in the public domain. As an ex-TV type, watermarks spring to mind as a way to prove ownership of digital content.

This debate isn’t new. The first computer programmer, Ada Lovelace, predicted the technological creation of art when she wrote about the idea of “poetical science” in the 1840s.

This was the same decade Paul Delaroche declared painting to be mort – but instead we saw the emergence of new genres of painting, such as impressionism. There was a shift in how people approached art.

Koach looks to the past for a hopeful take on the future: “AI art might shift our values away from, ‘Does this image have all the colours, compositions and styles that I want?’ to, ‘Is this image meaningful or special in some other way?’.”

But if we want to live with this technology, we need to be proactive. Learn about the complexity and bias within AI and keep a close eye on where it’s going. Regulate the industry, to protect users and creators. And outsmart the algorithms.

• Tracey Spicer’s book, Man-Made: How the bias of the past is being built into the future, is out May 2023

Source link