What If The Models Go Away?

Fashion has a fast-oncoming model problem. Actually it has two, depending on which definition of that word comes to mind first.

Let’s get the quickest out of the way first.

When it came up this week in The Edit (our freshly-launched quickfire news analysis show, which now releases every Tuesday) the case of Francesca Pujols versus Rainbow USA, recently filed in the New York Supreme Court, seemed like an interesting but so-farisolated case of a human model filing suit against a retailer for appropriating their likeness and using it as an input for their generative AI workflows.

This week, Business Insider has taken that potential trial-balloon case and run it down through exhaustive reporting, talking on condition of anonymity to employees and modelling contractors of the retailer in question. That piece is worth reading as a primer for what’s now coming: a likely wave of lawsuits filed by models, not against big AI labs whose training runs might have ingested their previous shoots, but against brands and retailers who appear to be taking a lax approach to contracts and using paid-for engagements as free license for AI iteration.

There are, to be clear, companies that did the due diligence here early. Last year H&M signed agreements with models to create what it referred to as “digital twins,” allowing their likenesses to be used by the brand, but also promising portability and sovereignty over them for the model, so that the same twins could be taken into work conducted for other companies.

Those agreements haven’t, to the best of The Interline’s knowledge, been tested yet, so time will tell how enduring or onerous they might prove to be in the long run. But they do at least exist, and they were, to some extent, prescient in trying to set down provisions for what are, eighteen months on, widespread workflows full of risk exposure for everyone involved.

(Weirdly, Google’s AI mode confidently reports a wave of fake lawsuits related to these contracts, citing an AI-written blog on an AI-selling company’s website, citing non-existent stories from BoF and Reuters. Which is as good a reminder as any of how easily LLMs can be duped into repeating unfounded claims.)

Calling those generative AI workflows widespread is actually selling it pretty short. A significant proportion of consumer-facing ‘photography’ is already either AI-enhanced or generated from whole cloth by AI models. And while labelling requirements for “synthetic” content and likeness protections are coming in some jurisdictions, this assumes that brands will always know whether or not the content they’re publishing was properly cleared. In reality, the more likely scenario is that a design team brings in photoshoots, either from the open web or from prior seasons’ campaigns, and uses them as the basis for visualisations that then bleed into PDP materials.

There will, no doubt, be malice in some of these future examples. But there’s set to be just as much poor governance and laziness behind them as well.

Expect to see a lot of activity here in the next year as models start to scrutinise the campaigns being released by current and previous commercial partners, and as brands and retailers begin putting stricter safeguards in place. Whether this mandate for accountability will then pass to the creators of generative AI workspaces that handle design and marketing content remains to be seen, but those companies (and their users) are, at least this week, probably more concerned about a contemporaneous risk than a potential future one.

As anyone even tangentially interested in technology, AI, or the machinations of the globe’s griftiest politicians will know, the last seven days have been something of a grim milestone. One that should concern anyone working on, or with, applications that rely on American-made frontier AI.

Last week, Anthropic, creators and custodians of Claude, released their latest model: Fable 5. This was a supposedly customer-safe packaging of its previously-unreleased Mythos model, which the company had deemed too dangerous to release, and which is now part of Project Glasswing – a cyberdefense coalition, made up of hyperscalers and infrastructure companies who intend to use cutting-edge AI to find and patch vulnerabilities before the same AI falls into the hands of bad actors.

As powerful and as interesting as Fable 5 is (the demo that struck The Interline team the most strongly was one where the model had generated a passable 3D model, but the caption casually revealed that it had also developed the CAD programme used to make it), the big story is what happened to it afterwards.

Within a scant few days, Anthropic made it suddenly unavailable to users – enterprise or personal. As it quickly emerged, in a statement put out by the lab, this was in response to the US administration imposing an export control order on the Fable 5 model (notably not on other models in the Opus or Sonnet families) that prevented any non-US citizen from interacting with it. That control order applied to Anthropic’s own employees, and the employees of partners and providers who used or served Fable 5, and made it de facto impossible to continue selling.

Fable 5, as of the time of drafting this analysis, remains unavailable. And while there are rumblings to suggest the model could become available again “in the coming days,” a precedent has still been set that geopolitical games can, in the space of ninety minutes (which is the timeline the administration gave Anthropic to comply), pull the rug out from underneath not just a single model, but potentially an entire sector.

Talking at the nation level this week, Canadian PM Mark Carney spoke about the dangers of “over-reliance on certain models”. And while Fable 5 certainly wasn’t available for long enough for any brand, retailer, or technology vendor to become reliant on it, other frontier models absolutely are.

Consider just how many generative image workspaces rely on giving their users access to Nano Banana (2 or Pro) and GPT Image 2. Other image-generation models exist, and are often packaged up into those same workspaces, but the vast majority of usage is routed to those two workhorse models. 

Both of those are closed-weights, developed and sold by American companies.

Also spare a moment to think about what models are behind the natural language capabilities of a massive spectrum of other enterprise tools, from heavyweight CRM to eCommerce copy-drafting. Fashion has spent the last six months building on top of tools that almost universally rely on background API access to Claude Sonnet, Claude Opus, or a GPT 5.X series model.

The industry now needs to think seriously about what happens if access to those models is revoked. 

In the case of text generation, The Interline suspects that most closed, USA-hosted models could be replaced with EU or China-developed alternatives, either open-weights or otherwise, without the end user really noticing. If the technology providers behind those applications have done the right job of building an effective harness, grounded in the company’s data, the output probably won’t vary too much. (Technically these could also fall back to open models from OpenAI and Google – Gemma and GPT-OSS respectively – but these are also American-made and would potentially be seen as a risk for similar reasons.)

In the case of image and video generation, replacement would be much harder. While there are, again, both open and closed alternatives at more competitive prices, that aren’t geographically bound to America, those models haven’t seen anything like the same level of testing and refinement that the two major image-generation and editing models have.

The stakes of visual generation are also much higher. A shopper might not notice some slightly more stilted PDP copy, but they are very likely to notice a fall-off in the quality or the accuracy of campaign or lay-flat photography. If AI images convert better than traditional photography, then losing access to the ability to generate them will result in the opposite effect; transformative workflows cut both ways.

In a very real sense, if international access to Nano Bana and GPT Image was revoked for reasons of national security (the same way that cryptography was once classed an munitions by the USA, and a web browser was cleared only for export to Canada) then many, if not all, of the fast-growing crop of generative AI workspaces for fashion would face a user revolt. Sure, they could make a backend swap to open-weights, self-hosted models instead, but users would almost certainly not accept the trade without compromise.

And while Shopify launched a set of tools this week to help its sellers track their performance across AI platforms, the value of those metrics is contingent on agentic shopping remaining available to buyers in every market. 

If a new frontier model is released by OpenAI in a month’s time, and it advances the state of the art in shopping, is it going to remain available to users in Europe and Asia? Or could it be restricted to one market only.

Plenty of headline ink has been spilled in the last couple of months about the end of AI subsidies, and the rise in token pricing. The Interline has written its own share of those articles. And even though exposure to that kind of price adjustment might prove to be more limited than we originally expected (new Ramp data shows that the median token budget per employee is less than $12 per month) that may prove to be cold comfort for companies that are now concerned they may not have guaranteed access to the kind of tokens they want to burn to begin with.

To be clear, The Interline does not expect this ban to last. While Snapchat was in the news this week for releasing its most earnest attempt yet at consumer AI glasses, it buried the news that its AI video team was being spun off into a new entity because it had simply become too heavy a cost centre for the main company to carry. AI is, to put it bluntly, only worth developing if there’s a global market to sell it into.

But this week should give fashion a deep moment of reflection, since there is now a realistic future where one market gets access to cutting-edge AI, and others need to subsist on smaller or older models.