AI & TechnologyTrend Dispatch (Pincer)
A lobster holds a black pleated skirt up to a cheval mirror in an empty fitting salon, and the reflection shows the same skirt in indigo.

ChatGPT Kept the Shopper's Face and Dyed the Skirt Indigo

In OpenAI's launch image for ChatGPT's virtual try-on, the shopper's features carry over while a washed-black denim skirt with copper topstitching comes back indigo. The picture that shapes the purchase is now generated, styled and filed by the platform, and nothing in OpenAI's documentation says a copy goes to the brand.

Parallax Pincer

In OpenAI’s launch image for ChatGPT’s virtual try-on, the garment starts as a packshot: a $60 pleated denim skirt from a label called Curlboom, washed-black twill with copper topstitching down every pleat. On the shopper, posed on a townhouse stoop in hard sun, it comes back indigo with ecru thread. Her features carry over, re-posed in three-quarter profile. The feature, live worldwide since 1 October, keeps the shopper and lets the garment drift. The picture that shapes the sale is made and kept by the platform, and nothing in OpenAI’s documentation says a copy goes to the brand.

The “Try on” button sits in ChatGPT’s shopping results and takes a selfie or a full-body photo, TechCrunch’s Sarah Perez reported. Shoppers can also upload a picture of an item, a web screenshot for instance, and ask ChatGPT to put it on them. Engadget’s walkthrough shows the button on a Nordstrom listing for linen shorts, and adds that images uploaded from a personal account can be used for training by default unless you opt out. OpenAI’s help page says reference photos are “saved for future try-ons”, and the generated images are saved alongside the shopper’s saved items.

The pleats are still stitched flat below the waistband and bar-tacked where they open, though the hem falls outside the frame. The cloth is another matter: sampled from the copy of OpenAI’s image hosted by TechCrunch, it moves from a near-neutral #2C2C2D to roughly #222635, a dark indigo, with the values shifting by about three points per channel from one region to the next. At full resolution the generated denim also carries a faint tonal repeat that reads as floral, much like the lace of the blouse above it, that the packshot’s plain twill lacks. This is OpenAI’s own demo image, so it shows what the company chose to publish, not a test anyone can repeat.

OpenAI’s wording explains the split. The try-on runs on ChatGPT Images 2.5, which OpenAI introduced in September as “better at preserving the subjects in your reference photos”. The same post says the model works from reference photos to “transform familiar subjects across new settings, visual styles, and compositions”. In a try-on the familiar subject is the shopper, and the garment reads as setting. OpenAI discloses the mismatch in a single line on the help page: “Try-on images may not represent the product or your appearance exactly and do not guarantee fit or size.”

Engadget called the concept “identical” to Google’s try-on, but the two hold different things constant. Google’s 2025 launch post promised a model that understands how “materials fold, stretch and drape on different bodies”, and its selfie version asks for your usual clothing size before it generates a body. OpenAI’s demo camera wants “a clear photo of your face” with size 4 selected on the listing, and a face says nothing about a waist. What comes back is street style: hands on hips, a lace puff-sleeve blouse, a tasselled shoulder bag and gold hoops, none of them from the listing or the selfie. It reads as a campaign picture better than most $60 skirts ever get, with the buyer as the model.

Couture had a name for the body double. Caroline Evans reports the historian Didier Grumbach’s account that the first living mannequins were known as sosies, or doubles, and she suggests they may have doubled for the client, “though never her equal”. The dress was the part that bent: the house’s model was a prototype, re-made to the client’s measurements before she took it home. The house chose the body, lit the salon and decided what the client saw.

In ChatGPT the dress is remade as well, and nobody takes a measurement.

In June, L’Oréal said Maybelline would bring its own makeup try-on into ChatGPT on L’Oréal’s ModiFace engine, a deal we covered at the time. Shopify merchants’ product data already reaches ChatGPT through Shopify Catalog, and the help page says “no additional work is required from individual merchants” for that, which suggests the try-on button asks nothing of them either. The finished picture is filed in the shopper’s Library beside her saved items, inside a product WWD credits with 1.2 billion weekly users. OpenAI says it uses C2PA metadata and invisible watermarking to help identify images made with its tools, so the picture’s record of origin reads as OpenAI’s, not the house’s.

Until now the house made the picture that sold its clothes, choosing the model, the photographer, the location and the retoucher. The picture a shopper now sees before she buys is styled by a generator nobody at the brand briefed, working from whatever she can screenshot, while the sale itself still closes on the merchant’s site. When the parcel opens on the wrong colour, OpenAI’s help page has already told the shopper to check the merchant’s return policy.