Does AI have as much taste as me?

Hi friend, I'm so glad you're here.

There's a new frontier in market research happening quietly inside consumer brands, and once you see it, you can't unsee it.

I came across a Business of Fashion piece on synthetic consumer research recently and couldn't stop thinking about it, so I went down the rabbit hole.

The promise

Imagine running a focus group in an afternoon instead of three months. Imagine testing a hundred product concepts before committing fabric to a single one. Imagine asking ten thousand "consumers" what they think of your fall collection, and getting answers by Friday.

That's the pitch behind synthetic consumer research: AI-generated personas that simulate how real shoppers would respond to products, campaigns, and concepts. Companies like Evidenza, Lakmoos, and Synthetic Users are raising serious money on this idea. Qualtrics (the giant of traditional survey research) launched its own synthetic panel last year and predicts that more than half of all market research will be done this way within three years.

The validation numbers are real, and that's what makes this hard to dismiss. A recent PyMC Labs study tested synthetic consumers against 9,300 real human responses across 57 surveys from a major consumer products company and hit 90% correlation on product rankings — with synthetic respondents showing less positivity bias than the human panels, meaning they discriminated more sharply between good and mediocre concepts. Dollar Shave Club (a direct-to-consumer razor subscription brand) used Qualtrics' synthetic panel to validate expansion into a new consumer segment and reportedly produced "identical thematic segmentation" to their previous human research, compressing weeks of work into days.

So ... It looks like it works, at least for some things.

How it actually works

The basic idea is simple. You feed a large language model a description of a target consumer (demographics, lifestyle, values and past behavior) and ask it to respond as that person. Multiply that by thousands of personas with different profiles, and you've got a synthetic panel. Ask them what they think of your new dress, your campaign concept, your pricing tier, and they answer.

The better platforms train their models on real survey archives, behavioral data, and customer reviews, so the synthetic responses are grounded in actual human patterns rather than the model's general guesses. Bellomy, a long-established US market research firm, calls this "digital twinning." Qualtrics built theirs on 25 years of its own research data.

In practice, a brand uses it like this: instead of recruiting a panel for early-stage concept testing, you run a synthetic round first to filter out the obvious losers, then commit your expensive human research budget to the concepts that survive.

Why fashion is paying attention

I'm partial to this topic because I spent a few years working on a related challenge: using predictive analytics to help fashion brands design, buy, and plan smarter. So when I see synthetic research arrive in fashion, I get genuinely interested ... and genuinely cautious.

Excited, because the stakes are enormous. Fashion is one of the most wasteful industries on the planet. Brands overproduce because they don't know what will sell. Trend cycles are punishing. Getting it wrong means markdowns, returns, and clothes that end up in landfills. If we could actually decode why a customer buys this dress in this color at this moment, we'd bring less to the market and waste less. Technology has to be part of that answer.

Cautious, because the formula for taste has eluded us for a reason.

Can taste even be quantified?

Researchers have been chipping away at this question for decades, and lately, machine learning has joined the effort.

The classical foundations are sociological. Bourdieu argued that taste is a form of cultural capital, a way we signal class, education, and identity through what we like. Simmel described fashion (essay here) as a constant tension between conformity and distinction: copying the people we want to belong to while trying to stand out from everyone else. In both views, taste isn't really about the object. It's about what the object says about you, to whom, and when.

Recent computational research has tried to operationalize this. A 2023 Cornell and Google project called Fashionpedia-Taste built a dataset of 10,000 annotations explaining why subjects liked or disliked specific fashion images. They broke each judgment into three layers: the localized attributes the person responded to (a neckline, a fabric, a silhouette), the visual attention (where on the garment the eye actually focused) and the caption, meaning the language the person used to describe what they saw ("vintage," "structured," "too try-hard"). The insight underneath the project is important: even when two people like the same dress, they often like it for completely different reasons, and a recommendation system that doesn't know which reason will eventually fail.

Other research treats taste as a moving target shaped by social signals. Trickle-down theories used to assume taste flowed from elite designers to mass consumers over seasons. On TikTok, the same flow happens in three days. Algorithms compress trend cycles by rewarding engagement, which means a niche aesthetic can become saturated and exhausted before a buying team has finalized next spring's order.

The honest summary is this: serious people are making serious progress on quantifying parts of taste. But nobody has cracked it, and there's a structural reason why — which is exactly where synthetic consumers run into trouble.

The catch

Large language models work by predicting the most probable next thing based on patterns in their training data. That's their superpower and their limitation. A synthetic consumer is, by design, an averaging machine. It tells you what a person like this would probably say, based on what people like that have said before.

That's useful for a lot of things. It's terrible for fashion's most important question: what's next?

A synthetic panel trained on past data can tell you that your new leopard print sneaker resemble things people previously liked. It cannot tell you whether the cultural moment has shifted away from prints altogether while you weren't looking. It cannot tell you that the exact same shoe will feel fresh in London and dated in Seoul. It cannot tell you that the influencer who would have made it cool just got cancelled for a 15-year-old tweet!

And there's a deeper issue. A recent systematic review of 50 studies on how consumers actually respond to AI in fashion found that consumers contradict themselves constantly. They say they want personalization but resist surveillance. They appreciate AI recommendations but find them less authentic than human ones. They like human-like AI until it crosses a line, and then they hate it. If real consumers are this internally inconsistent, what exactly is a synthetic consumer modeling? You're training AI on the expressed preferences of humans whose actual behavior contradicts what they say.

My take

Synthetic consumer research is probably very good at de-risking the safe middle of fashion. Basics (The Uniqlo, COS & co ...). Color refreshes. Broad campaign messaging. Pricing tiers. The stuff where consumer behavior is rationalizable and the question is "will enough people buy this" rather than "is this culturally interesting."

It's probably very bad at the thing fashion actually runs on: novelty, taste-making, the feeling that something is now. The part of fashion that creates demand instead of predicting it.

And here's the uncomfortable question for the industry: if everyone uses synthetic research to de-risk the middle, do we accelerate the homogenization people already complain about? Fashion is already drifting toward sameness, the same minimalist Instagram aesthetic, the same neutral palettes, the same algorithm-friendly silhouettes. Synthetic research, used uncritically, would pour gasoline on that fire.

So is synthetic research the next fashion tech frontier?

Honestly? No. Not as a replacement, anyway.

I think it's a useful complement to human research, a way to filter early and iterate fast but the brands treating it as a substitute are going to learn an expensive lesson. Synthetic research can probably do real work in industries where purchases are rationalizable and quantifiable. Fashion isn't one of them.

Fashion is a distinctive space because fashion consumption is tied to self-expression, aesthetic judgment, social meaning, and emotion. I buy when I'm happy and I buy when I'm sad. I treat myself when I accomplish something, and I shop for comfort when I haven't. I buy to feel good about myself and I buy to belong. I buy to tell the world who I am and sometimes, even who I'd like to be.

Good luck modeling that, lol.

The road to decoding the formula of fashion purchase is still long, and honestly, that's why I love this industry. The mystery is the point.

What do you think? I'd love to hear other perspectives.

— Lauren

Fashion Stack is a personal project. Views are my own and do not represent those of any employer.