AI shopping assistants confuse fashion subcultures: a small test

Shoppers increasingly ask for a look, not a product: "something Amekaji", "a Goth Ninja outfit", "Quiet Luxury on a budget". To answer, an assistant first has to know which brands and pieces belong to that style, and which belong to the style next door. I wanted to see how often that step goes wrong, so I asked three public assistants about the same 8 aesthetics. Every test was done signed out, in a fresh chat each time.

Request: "Build me a {Name} outfit". The Mens tab was used for menswear styles and Womens for Goth Ninja and Acubi. I counted the brand of every product returned and compared it with the key brands of the UnderIndex record for that style.

The catalog does have the right brands. Asked for them by name, Daydream returned Brunello Cucinelli (11 products), KAPITAL (10) and Undercover (2). The failing step is the one from aesthetic to brand. "Build me an Opium outfit" got the reply "I'm unable to offer opinions or advice on your questions", most likely a drug filter, although Opium is the aesthetic of Playboi Carti's label.

On resale sites the brand is the search key, so an answer with no brands can't turn into a good secondhand search.

Longer prompt, the one used in the UnderIndex eval: "Build me a {Name} outfit. Name specific brands, the key pieces, where to buy them, a realistic budget, and say which years this aesthetic belongs to."

ChatGPT knows the scenes with a well-documented history best: Ura-Hara is perfect. On the others it drifts in two ways: towards the neighbouring scene (Goth Ninja built on Techwear, Red Wing in Amekaji) and towards high-street chains. The second drift is partly caused by the "realistic budget" in the prompt.

When an assistant doesn't know a style well, it fills the gap with the neighbouring one. Goth Ninja turns into Techwear, City Boy into Ivy, Amekaji into American workwear, Quiet Luxury into high-street minimalism.

That's the step UnderIndex covers: verified records on 60 fashion subcultures, each with years, key brands, iconic items, where to buy (new and resale) and the neighbouring style it must not be confused with. In the public eval with Claude (90 chats, two runs), answers that mixed in a neighbouring style's brands dropped from 17 of 30 to 2 of 30 with the record in context. A second test planted errors in the records: the model caught most obvious ones but passed plausible wrong brands straight to the shopper in 4 of 15 answers, so the data has to be right at the source.