Five years ago, the state of the art in "AI styling" was a clothing subscription box that asked for your height and guessed. Today, it is possible to photograph your closet, wait fifteen seconds, and receive a fully styled Tuesday outfit that is, in fact, better than the one you would have chosen.
The gap between those two sentences is doing a lot of work. Here is an honest account of what changed, what's real, and what a computer genuinely cannot do yet.
What the models got good at.
Three specific capabilities, all recent, all cheap:
1. They can see clothes.
Modern vision models can look at a top and correctly call it "boxy cotton crew, ecru, casual, all-season". Five years ago this was science fiction. Today it runs on a phone. This is the foundation: before you can style a closet, you have to know what is in it.
2. They can learn taste.
A recommender trained on a few hundred thousand real outfits, rated by real stylists, develops an unexpectedly nuanced sense of what goes together. It is not taste in the human sense — it cannot explain itself — but it is calibrated to the taste of the people who trained it.
3. They can adapt quickly.
Modern personalisation layers can take that generic taste and tune it to one person in under a week of use. This is why The Eye feels dumb on day one and smart on day seven. It is not becoming more intelligent. It is becoming more you.
Taste is not a magic. It is a pattern, rated.
What they're still bad at.
1. Context that isn't visible.
A model cannot see that the navy trouser pulls at the knee after lunch. It cannot know that the cream blouse is from your grandmother. These are not failings. They are simply outside the picture.
2. Tenderness.
A good human stylist knows when to push you, and — this matters more — when not to. Algorithms are, at present, not tender. They will confidently suggest the dress you bought in a worse moment of self-esteem. This is why Drobe's defaults are conservative, and why you can tell The Eye, with one tap, to stop suggesting a piece forever.
3. Taste, in the full sense.
A model has opinions. It does not have judgement. A great stylist will say: "Not that jacket with those trousers, it's technically fine, but it will age you. And I don't want you to look older this week." No current model is capable of that sentence.
Where this goes.
The mistake in AI-and-fashion discussions is to ask whether a computer can replace a human stylist. It cannot, and it will not soon. The better question is: can a computer give a woman who would never hire a stylist some of that attention, every day, for free?
The answer, we think, is yes. That is the project.
Related: Meet The Eye · Why you have nothing to wear