Describe the copyright exposure of a generative feature.

Selvage Studio builds custom garments to measure. StyleCast is the feature that takes a customer's photo and a style reference and renders a mockup of the finished piece before a single stitch is cut. Coen Vermaas is the product lead who has to put a real number on the copyright exposure for the executive team, not just a feeling about it.

How to answer this, stage by stage

Nobody is grading whether your number lands exactly right. They're grading whether you can show the arithmetic and name the one input worth fixing first.

Let's learn

StyleCast reads a customer's photo and a chosen style reference, then renders a mockup of the finished garment, so a customer can see roughly what they're ordering before any fabric is cut.

Before StyleCast shipped, Selvage's design team built training images the fast, free way: they scraped publicly posted runway photography, since it was abundant, current, and nobody stopped them.

Selvage's design team treated these two as the same thing. They are not.

Now, with StyleCast live, every mockup blends a customer's own photo with patterns the model learned from that scraped corpus, including silhouettes lifted from specific, still-in-copyright designer collections.

More than half the corpus came from the one source with no confirmed right to reuse it this way.

The turn: the exposure isn't really about how many mockups get generated. It's about how many of those mockups land close enough to a specific, identifiable, still-in-copyright design that a designer could reasonably say "that's mine."

What I would leave alone: the 10 percent of the corpus made up of customers' own uploaded photos needs no rework at all. Those images came with the customer's own consent for their own garment, and reviewing them the way we're reviewing the runway slice would waste effort on the safest part of the corpus.

The lesson: a copyright exposure estimate isn't a legal exercise bolted onto a finished feature. It's a direct readout of one design decision, made back when the training corpus was built, about which shortcuts felt harmless at the time.

Now here is the same thing as a story

The short version above is what you'd say defending this estimate to Selvage's board. Read this one for how the near miss actually happened.

Coen Vermaas has led Selvage's product team for four years, mostly working on fit prediction and sizing, before StyleCast became his to own as well.

StyleCast launched to good reviews. Customers loved seeing a rendered mockup before committing to a custom order, and the design team kept feeding the model fresh runway photography every season, since it was the easiest way to keep the style references current.

The largest slice of the corpus sat in exactly the corner with the least clarity about who actually owned it.

Then, three months in, a customer service rep forwarded Coen a mockup a customer had shared on social media, tagging it with a designer's name and the caption "wait, did they just copy this dress?"

Side by side, the mockup and the original runway gown were close enough that even a customer noticed without being told to look.

Nobody had built a step to check a generated mockup against recent, identifiable designs before it reached a customer. The pipeline went straight from generation to display, with no gate in between.

Four steps ran the way they were supposed to. The missing fifth step was the one that would have caught this before a customer ever saw it.

Coen pulled the design team together and sorted every training image by where it actually came from, for the first time treating "licensed," "scraped," and "customer-consented" as three separate categories instead of one undifferentiated pile.

Sorting the corpus this way turned an abstract worry into a concrete removal list.

Legal helped write new contract language while the removal list was still being built, so customers ordering during the cleanup would already be covered.

Four plain terms, written before the next mockup shipped, not after a designer's lawyer called.

The old corpus treated any publicly visible photo as fair game the moment it could be downloaded. The new one asks a specific rights question about every image before it ever trains anything.

I approved scraping runway photography because it was the fastest way to keep the style library current, and at the time nothing about the model felt customer-facing enough to matter. It took one tagged social post, from a customer who never even knew there was a legal question at stake, to see that "current" and "cleared" had never been the same requirement.

BOUND, the exposure broken openNot a single guess. BOUND is what forces every dollar in that number to say where it came from.

The recap, one line per letter: break it down is claims per year times cost per claim, own numbers is one to four claims at 150 to 400 thousand each, use a range is the 150,000 to 1.6 million spread, nail the sanity check is the high case against a 500,000 dollar insurance cap, and direction is the scraped-photo share as the number worth fixing first.

And if you want to be sure it really works, try it somewhere elseSame five letters, a home-renovation visualizer instead of a garment mockup. This time the exposure is measured in complaints, not dollars.

Birchcroft Renovations uses an AI feature that renders a homeowner's kitchen with a new layout, pulled partly from photos of finished projects posted by other contractors and design blogs. Wren Halloway runs product there, and the same estimation question looks different once you're not talking about designer gowns.

Break it down: expected exposure equals expected takedown demands per year from contractors whose finished work appears recognizably in a rendered mockup, times the cost of legal review and response per demand, since litigation here is rare but demand letters are common. Own numbers: an estimated two to six demand letters a year, given how much of the training set comes from contractor portfolio sites with ambiguous reuse terms, at roughly 3,000 to 8,000 dollars in legal review time per letter. Use a range: 6,000 dollars low case, up to 48,000 dollars high case, far smaller in absolute terms than Selvage's exposure, but a real cost against a much thinner renovation-industry margin. Nail the sanity check: the high case is still less than one month of the legal team's retainer, so it's real but not yet a board-level number. Direction: the single biggest lever is tagging which portfolio photos came with an explicit reuse permission versus which were simply scraped from a public site, the same underlying assumption as Selvage's, wearing a different industry's numbers.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "150,000 to 1.6 million a year, and the scraped-photo share is what swings it," and stop.
Cost: there's no budget this quarter to relicense the whole corpus at once. Say so honestly, and start by relicensing just the most recent two seasons, since recent designs are the ones most likely to still be actively enforced.
The model gets better, for real: if StyleCast's rendering quality improves, that doesn't lower the exposure, it raises it. A sharper, more accurate mockup is more likely to look recognizably close to the original design, not less.

Where people run it wrong.
They treat "publicly visible" and "free to use" as the same fact, when they are legally unrelated.
They quote a single exposure number with no range, which hides how much the whole estimate leans on one unverified assumption.
They wait for a claim to arrive before building the similarity check that should have been part of the pipeline from day one.

How to use it live. When asked to describe copyright exposure, don't start with the legal theory. Start with where the training data actually came from, and let the arithmetic follow from that.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0