Generative AI’s Big ‘Flaw’ May Also Be Its Superpower

There are a number of differences between lawyers and fashion designers, but a key one is that lawyers generally aren’t supposed to fabricate things. It’s a lesson one New York attorney learned the hard way when he was sanctioned earlier this year after the court brief he wrote with generative artificial intelligence included several made-up cases.

But in conversations with designers, several have said the technology’s disregard for considerations like standard clothing construction or basic physics, as well as its vulnerability to “hallucinations,” may be its most powerful asset.

Generative AI models don’t understand the meaning of the text or images they create. They’re predicting how words should appear together in a sentence or how pixels should join to form a picture based on patterns observed in the giant amounts of data they’ve ingested. But they can make mistakes or get the general idea right but the specifics wrong. That tendency is a major drawback when using the technology for factual purposes, such as answering medical questions or writing legal briefs.

In design, however, there generally isn’t a binary between right and wrong. In fact, the technology’s vulnerability to getting things “wrong” can yield unexpected new ideas. It’s probably also the greatest challenge of using it, as a creative team then needs to turn that output into a real-world product.

“Where it’s just completely unlimited by any sort of reality, [that’s where] I could see AI help the design process,” said Julius Juul, co-founder and creative director of Copenhagen-based Heliot Emil, which, with the help of an outside agency, used generative AI in the creation of its Spring/Summer 2024 collection. “That’s where it gets interesting.”

Juul gave the example of a chain necklace generated with AI, except it was only half a necklace. The AI didn’t understand that it would just fall off the wearer in the real world. To replicate it physically would require some trickery, perhaps a nearly invisible thread completing the loop for instance. But even ideas capable of being turned into physical garments may be difficult to translate. It took Collina Strada four weeks to figure out how to perfectly construct the sleeve on one AI-generated dress.

Still, Juul said the technology yielded ideas he might have never thought of, including less gravity-defying examples like some unusual silhouettes that featured in the brand’s recent collection.

“For the design process, what you could use it for is to push yourself outside the constraints that you sometimes put up in your mind, unintentionally or subconsciously,” he said.

0 of 8

Though not specifically about using AI for imagery or design, a recent Boston Consulting Group study highlighted the split in generative AI’s value for creative purposes versus other tasks. Enlisting more than 750 of the firm’s consultants around the world as subjects, it had them use GPT-4 — the most-advanced, publicly available version of the AI language model underlying ChatGPT — for two different types of work. The first, which it called creative innovation, focused on coming up with new products and devising go-to-market plans. The second was business problem solving, which entailed participants spotting the cause of a company’s challenges using executive interviews and performance data.

BCG found those using AI for creative innovation performed about 40 percent better than a control group that didn’t use it. Conversely, participants who used AI for business problem solving performed 23 percent worse than the control group. “And even participants who were warned about the possibility of wrong answers from the tool did not challenge its output,” the authors noted.

It wasn’t all good news on the creative front. One big shortcoming the study found was that the generative-AI users offered less diversity in their ideas, since they tended to get similar responses from the AI.

As the technology continues to improve and developers learn to minimise its hallucinations, designers could also find that AI gets better at producing realistic garments — ones without the odd flourishes that can currently inspire designers.

Fashion brands trying to stand out from the crowd may find they need to work a bit harder to get the best output from the technology, although companies in the business of churning out mass quantities of garments may be content to have the AI produce fairly conventional designs; in fact, they could prefer it if they’re just trying to quickly generate new variations on existing concepts — one of generative AI’s potential strengths. But brands that centre their creativity and want to maintain a distinctive style won’t want the same looks as everyone else.

Heliot Emil used its previous collections and imagery to train the AI so its output would be based on the brand’s visual DNA, a process much like that used by Collina Strada. That imagery provided the basis for physical garments, which featured in its September runway show. The brand also fed pictures of those garments back into the AI to create novel digital designs, which it blended in alongside the photos of its physical pieces on its website.

The results all bear a consistent style blending structured tailoring and chaotic draping, with fabrics wrapping the body or dangling in tendrils around it. The big giveaway of the purely AI looks is that, on close examination, you can see they would probably be impossible to construct as they are.

Juul said that, while the strange quirks of the AI were helpful to his design process, they also mean it’s likely to be some time before it’s capable of taking over the role of a human designer. Companies such as CALA and Raspberry.ai are trying to streamline the process of using AI for design through interfaces and other capabilities to specialise it for creating manufacturable garments. But it still requires a human to identify which designs are new and interesting enough to be worth producing.