The water can be dangerous. We can easily drown in water. But he has no intention. And the challenge humans face when it comes to water is learning to swim, build boats and dams, and find ways to wield its power.
“You can take two pictures, and that’s great, but you take 100,000 pictures and you have a real physical feeling of drowning,” Holz says in an interview with Fortune. “So we’re trying to understand how people are taught to swim? And how do you build these boats that allow them to sail and be empowered and sort of navigate the ocean of imagination, instead of drowning?
AI image generators have proliferated across Silicon Valley and gone viral on social media. Just several weeks ago, it became nearly impossible to scroll Instagram without seeing Lensa AI’s “magic avatars,” which are colorful digital selfies made with an AI-powered editing app.
“Over the last 12 months, the development of these technologies has been quite immense,” says Mhairi Aitken, research ethicist at the Alan Turing Institute, the UK’s national institute for data science and artificial intelligence. [A.I. image generators] generate a certain output without necessarily having to understand what is the process for which it was created or the technology behind it.
Lensa AI is an all-in-one image editing app for selfies and photo editing. Courtesy of Lensa AI
The models behind these AI image generators are permeating smartphones as recent discoveries delve into the models’ ability to understand language and also create more realistic photos. “You’re teaching the system to familiarize itself with a lot of elements of the world,” Holz explains.
As a result, virtually any user can design, process and rework their own facial features in images uploaded to apps such as Lensa AI, which launched late last year and already has more than a million subscribers. Going forward, Lensa says it is looking to evolve the model into a one-stop-shop that can cater to all users’ needs around visual content creation and photography.
Art generated by artificial intelligence initially emerged in the 1960s, but many of the models used today are in their infancy. Midjourney, DALL E 2, and Imagen, some of the best-known players in the space, all debuted in 2022. Some of the biggest tech behemoths in the world are paying close attention. Google’s text-to-image beta AI model is Imagen, while there are reports that Microsoft is brooding a investment of 10 billion dollars in OpenAI, whose models include the ChatGPT chatbot and DALL E 2.
“These are some of the largest and most complicated AI models ever implemented in a consumer way,” Holz says. “It’s the first time that an ordinary person has come into contact with these huge and complex new models of artificial intelligence, which will define the next decade.”
But the new technology is also raising ethical questions about potential online harassment, deepfakes, consent, the hypersexualization of women, copyright, and the job security of visual artists.
Holz acknowledges that AI image generators, as is the case with most new technological advances, contain a lot of male bias. The humans behind these models still have work to do to understand the rules behind AI image generation, and more women should play a decisive role in the evolution of this technology.
At Midjourney, there was discussion about whether the lab should allow users to upload sexualized images. Take the example of a woman wearing a bikini on the beach. Should it be allowed? Midjourney brought together a group of women to ultimately decide that yes, the community could create bikini images, but that those images would be private to the user and not shared system-wide.
“I didn’t want to hear just one guy’s opinion on this,” Holz says. Specific phrases are blocked by Midjourney to prevent malicious images from proliferating within the system.
“Midjourney is trying to be a safe space for a wide variety of ages and all genders,” says Holz. “We are definitely more Disney of space”.
On the one hand, one could make a bleak argument that AI image generators, which again have no human intent, are simply reflecting our society back to us. But Aitken says it’s not good enough. “It shouldn’t just be a matter of taking the available data and saying, ‘That’s the way it is,'” says Aitken. “We are making choices about the data and experiences that are represented.”
Aitken adds that “we need to think more about representation within the tech industry. And we can ensure greater diversity within those processes, because it often happens that when biases emerge in datasets, it’s because they weren’t anticipated in the design or development process.”
Concerns about how these templates can be used for harassment, the promotion of bias or the creation of harmful images have led to calls for more guardrails. Google research shows some mixed opinions on the social impact of text-in-image generation. Those concerns were big enough that the tech giant decided against it publicly release the Imagen code or demo. Governments may also need to step in with regulation. China’s Cyberspace Administration of China has a new law that went into effect in January which requires AI images to be watermarked and the consent of individuals before they are deepfaked.
Visual artists have also expressed concern about how this new technology infringes on their rights or could even take away work they were previously paid for. The San Francisco Ballet recently experimented with Midjourney technology while creating an AI digital image for their production The Nutcracker. Users he flooded the social media post on Instagram with complaints.
In January, a bunch of AI image generators, including Midjourney, were named in a lawsuit who claimed that the dataset used for their products was trained on “billions of copyrighted images” and downloaded and used without compensation or consent from the artists. The lawsuit concerns violations of California’s unfair competition laws and the protection of artists’ intellectual property similar to what happened when streaming music technology emerged. The lawsuit was later dropped FortuneMidjourney’s interview and the publication has contacted Midjourney for further comment.
Holz says most of the people who use Midjourney aren’t artists, and very few people sell images made with the model.
“It’s almost as if the word AI is toxic, because we implicitly assume it’s here to replace and kill us,” Holz says. “An important thing is to understand how to improve people, rather than how to replace people.”