What an AI Stylist Cannot Do

House of Nova builds an AI stylist, which makes this an awkward article to write and the reason it is worth writing. Every claim below cuts against our own interest, and you should weight it accordingly.

It cannot see how something fits you

This is the hard limit, and it is not close to being solved.

Fit is a three-dimensional relationship between a garment's pattern and a particular body: where a shoulder seam lands, whether an armhole binds when you reach, how a hem falls on your specific leg length, how the fabric behaves when you sit. A model trained on catalogue photography has seen how clothes fall on sample bodies photographed to flatter them. It has not seen how they fall on you.

An AI stylist can tell you what a garment is cut to do, and if you tell it your proportions it can reason usefully about whether that cut tends to suit them. That is a genuinely different thing from knowing it fits, and any tool that blurs the two is overselling.

The practical consequence: use it to narrow a shortlist, not to skip a fitting. This is why returns are the honest measure of a styling tool, not engagement.

It does not know what suits you until you tell it

There is no way to infer taste from nothing. A stylist — human or otherwise — that has just met you is working from population averages until you supply the specifics, and population averages are exactly the recommendations that feel generic.

What actually improves the output is unglamorous: your proportions, the colours you already reach for, where you live and what the weather does, what you do for work, what you own. Every one of those is something you have to provide.

Be suspicious of a tool that claims to know your style from a single photograph or three quiz answers. It is producing something plausible, and plausible is not the same as personal.

It has no taste of its own

A model does not find something beautiful. It has learned what tends to be described as beautiful in the text and images it was trained on, and it reproduces the centre of that distribution.

This makes it reliable at conventional good judgement and structurally bad at the interesting edge. It will not put two things together that nobody has yet — it will put together the things that have been put together most often, expressed fluently. For most people on most days that is genuinely useful. It is not the same as a stylist with a point of view, and if what you want is the unexpected, this is not yet the tool for it.

It will be confidently wrong

Language models produce fluent output whether or not the underlying reasoning holds. The failure mode is not gibberish; it is a well-argued recommendation built on a wrong assumption — that a fabric is heavier than it is, that an occasion is more formal than it is, that a piece is in stock when it is not.

The tell is that the confidence never varies. A human stylist hedges when unsure. A model's certainty is not calibrated to its accuracy, so you have to supply the doubt yourself. Ask why. A recommendation that can explain itself in terms you can check is one you can evaluate; one that cannot is a guess in a nice sentence.

What it is genuinely good for

Having said all that, the things it does well are not trivial.

Narrowing. Going from a catalogue of thousands to six things worth looking at, filtered by an occasion you described in a sentence rather than by clicking through filters.

Stating the reasoning. A recommendation that comes with its argument attached — this shape because of that proportion, this colour against that palette — is something you can agree or disagree with. That is more useful than a ranked list.

Remembering. It does not forget what you said last time, and it does not get bored on the fifth question about the same event.

Being available. Most people have no access to a stylist at all. The comparison that matters is not against a good human stylist; it is against having no help, which is what most people currently have.

How to judge one

Ask it something it should refuse. A good tool will tell you it cannot confirm fit without measurements, or that it does not know whether something is in stock. A tool that answers everything with equal confidence has been built to impress rather than to be right, and it will eventually cost you a return, a bad purchase, or an occasion you were dressed wrong for.