The shock and awe of state-sponsored women’s fashion

20.10, Tuesday 16 Aug 2022

I think that, because we’re a capitalist society, we think of AIs as amplifiers for production and consumption. But they can force-multiply on any vector if suitably directed.

And, I don’t know, could you weaponise the fearsome AI that is the Gen Z fashion app Shein?


Back in 2017, Anna Batista asked (at Irenebrination): Can the Algorithm Become a Cool Fashion Designer?

Developed by Amazon’s San Francisco-based Lab126 - the company’s research and development hub - the algorithm uses a tool called generative adversarial network (GAN). … In a nutshell, the algorithm may spot a trend on Instagram, Pinterest, Facebook, or in its own collection images generated by Amazon’s Echo Look camera, and come up with new styles.

AI fashion.

A GAN is actually two AIs, a generator and a discriminator.

The generator sits there pumping out new dresses (or whatever). The discriminator does its best to recognises the dresses (or whatever) and score them. The generator learns how to improve its score. Ta-da, amazing dresses.

Though I don’t recall it taking over Amazon.


Based in China and shipping across 220 countries, Shein is the world’s largest fashion retailer, as of 2022 (Wikipedia).

The generator:

It starts with algorithmically scouring the internet and Shein’s own data to pull out fashion trends. As one of Google’s largest China-based customers, Shein has access to Google’s Trend Finder product, which allows for real-time granular tracking of clothing related search terms across various countries. This allowed Shein, for example, to accurately predict the popularity of lace in America during the summer of 2018. Combine that with Shein’s huge volume of 1st party data through its app from around the globe and software-human teams that scour competitors’ sites, and Shein understands what clothes consumers want now better than anyone with the possible exception of Amazon.

Shein feeds that data to its massive in-house design and prototyping team who can get a product from drawing board to production and live-online in as little as three-days.

From there, it can start with incredibly small batches, around as small as 10 items.

(Go read that entire breakdown of Shein’s business. The ERP innovation is remarkable.)

The discriminator:

Products in very small numbers are added to the app, and then clicks, views, purchases, and shares are monitored - automatically scaling orders to the network of factories. At great scale.

Shein churns out and tests thousands of different items simultaneously. Between July and December of 2021, it added anywhere between 2,000 and 10,000 individual styles to its app each day, according to data collected in the course of Rest of World’s investigations. The company confirmed that it starts by ordering a small batch of each garment, often a few dozen pieces, and then waits to see how buyers respond. If the cropped sweater vest is a hit, Shein orders more. It calls the system a “large-scale automated test and re-order (LATR) model”.

I mean.

While the bit of Shein that surfaces trends makes use of AI, it isn’t an AI designer in itself (the team of human designers was already 800 strong by 2016).

But throw consumerism into the loop, with feedback into ordering, and the entire thing resembles a giant Generative Adversarial Network, an AI for producing fashion lashed together out of software, supply chains, designers, and desire, teaching itself how to improve all the time.

Likewise I wouldn’t call the Facebook newsfeed algorithm an AI, but coupled with user clicks and eyeballs as a discriminator, I most definitely would. The trick is to include the human response.

(If we could talk to the Shein AI, I wonder what it would say? It would be like trying to talk to an intelligence emergent from the fluid dynamical storms of Jupiter. I wonder how we could send it a message, and if we would recognise any response?)

It’s tempting to think of this giant fashion GAN as neutral somehow. Like: it generates, we discriminate, and what comes out is fashion.

But it’s trivially possible to reach inside the machine. The prince of Shein could decide that everyone is going to wear blue next month, and could choose to only generate blue garments, and of course the thresher of consumerism can only discriminate over what it’s given…

So at that point the AI would grasp the flywheel, and blindly optimise its way to figuring out exactly how to make blue garments appealing and profitable.

I mean, AI is a fearsomely powerful gradient climber. It’s a weapon.

And it occurs to me that:

  • Shein is China-owned, and GANs its fashion into the world
  • TikTok is China-owned, and can boost trends, and suggests filters for how we look
  • Zoom’s software is developed in China, serves 3.3 trillion meeting minutes annually, and could - if it wanted - subtly and silently change my appearance, like my shirt colour or my weight or the percentage of eye contact, and you would never know.

I’m not picking on China especially here or suggesting they are actually up to anything or would have a motivation to do so, but this combination of appearance mediators makes me ask:

What would a state-sponsored fashion hack look like?

A fashion hack isn’t like the other global infrastructure exploits I’ve previously wondered about because it isn’t entirely obvious what you’d use it for. But the thing about state-sponsored attacks is that they’re a bit like magic tricks: they operate at a scale which is absurd, which makes them unimaginable, and that’s why they work. Like artificial weather or guided influenza.

It’s absurd to contemplate that plausibly deniable rainstorms might be directed to disrupt a UN weapons inspection team, or people peaking on the infectivity curve with flu are quietly standing next to diplomats 48 hours ahead of an important negotiation - but I bet it has happened.

Could a malevolent state actor

  • tweak the timings and appearances in online meetings to make the exec team of a competitor company simply… not like each other as much, or never have an effective conversation? (Could you measure that in the stock price, against corporate meeting software contracts?)
  • deliberately diverge the clothes preferences of two friendly nations, to undermine the common feeling they previously experienced?
  • associate a particular look with transcendant beauty, and deny sales that would create that look to particular nations, immediately creating a status gradient? (I mean, this is no different from how pop music and abstract expressionism were deployed in the 1960s by the West, only with new technology.)
  • get people to wear stupid hats, simply for their own amusement?


Hey, follow-up question: if we were in the middle of a giant fashion hack, how would we know?

If I got to pull the levers, I’d use Shein and TikTok to target my own civilians, and I would artificially boost acceptance and desire for cybernetically enhanced clothing, catalysing an arms race for cyborg prostheses providing the wearer with both superhuman powers (such as strength, speed, and musical accomplishment) and astounding aesthetics.

But that’s just me.

More posts tagged:
Follow-up posts:

If you enjoyed this post, please consider sharing it by email or on social media. Here’s the link. Thanks, —Matt.