If you’re cute on Instagram, it will sell. — Simon Porte Jacquemus.
There is a question no one in the fashion industry wants to answer out loud: who is actually making the clothes? Not the designer whose name is stitched into the lining. Not the studio team in some Parisian atelier. The honest answer, if you follow the decisions back far enough, is that a machine learning model running inside a server in Menlo Park has more influence over what ends up on a runway than most creative directors will ever admit.
This is not a conspiracy. It is not even subtle. The data has been public for years, the confessions have been made in glossy profiles and investor decks, and still the industry moves forward as though fashion remains a discipline of instinct and vision, the lone genius staring at a bolt of fabric and hearing something. That mythology is useful. It protects the premium. But it is increasingly fictional.
Simon Porte Jacquemus said it himself, without apology, in a 2019 profile for W Magazine. He designed the Le Chiquito handbag to be absurd, comically small, visually loud. He designed La Bomba, the cartoonishly oversized sun hat, because he understood that something that large would be photographed. His team told him no one would wear the hats. They sold hundreds. His own explanation for why he made both objects: if it is cute on Instagram, it will sell. He did not frame this as compromise. He framed it as strategy. The distinction matters.
What Jacquemus articulated so plainly, thousands of other designers are living without saying. The question of what gets made in fashion is no longer answered primarily in a studio. It is answered by what an algorithm surfaces, what a feed rewards, what a save or a share signals back to a recommendation engine that is, in its own mechanical way,expressing a preference. The algorithm has opinions. And the industry is listening.
To understand how this happened, you have to understand what Instagram’s recommendation system actually does. It does not simply show people content from accounts they follow. It predicts, with considerable accuracy, what a given user will engage with based on prior behavior, and it distributes content accordingly. What the platform rewards is retention: watch time on Reels, saves on carousels, DM shares, profile visits. These signals tell the algorithm that a piece of content has value, and the algorithm responds by amplifying it. The content that performs rises. The content that does not, disappears.
For a fashion brand, this creates a feedback loop with serious structural consequences. You post a look. The algorithm measures engagement. High engagement means the algorithm shows the look to more people, who engage with it, who buy it, who post their own versions, whose posts the algorithm then amplifies, which tells the brand what its customers want more of. This is not trend forecasting. This is trend manufacture, with the platform acting as the invisible production house.
Companies like Heuritech have made this feedback loop into a business. Their AI scans three million social media images every day, analyzing over two thousand fashion attributes including colors, silhouettes, prints, textures, and proportions. They claim predictive accuracy above ninety percent for up to two years in advance. Louis Vuitton uses it. Dior uses it. The practical implication of this technology is that a designer sitting down to plan a collection is increasingly doing so with a data report in hand that tells them which trends have momentum on social media, which are declining, and which the algorithm is likely to reward in the coming season. The creative brief has an analytics annex.
Stitch Fix’s vice president of buying described the shift plainly to NPR in 2025. Her team used AI to decide between a red stripe and a blue stripe on a shirt. Not as a tiebreaker. As the primary input. In the past, she said, you either made a gut decision or you waited weeks for overseas samples. Now you run the data. The intuition has not disappeared entirely, but it has moved to a secondary position, consulted after the machine has already spoken.
