Fashion Brands Using AI Models: The Verified Record

Fashion brands using AI models fall into two groups with opposite economics. High-volume retailers mostly work from replicas or AI edits of real models, then run them at catalog scale and report the numbers publicly. Brands chasing a campaign moment generate a fully synthetic face instead. The first group produced measurable production gains. The second produced headlines.

Almost every article on this topic lists the same eight benefits and names the same three brands. Very few check which brands did what, and the ones that check find the pattern reverses.

This guide works from company filings, corporate press releases, court reporting and Tier-1 news coverage, each with a date. Where a widely repeated claim has no source behind it, that is stated too.

Key Takeaways

• The volume operators did not buy synthetic models. Zalando, H&M and Zara built their programmes on replicas or AI edits of real models, and H&M and Zara say, or models report, that those models are paid. ASOS has not said how its model imagery is made.

• The best-evidenced number comes from a public filing. ASOS reported that AI Studios delivered more than 1,600 additional assets for one April campaign, work that would have needed around nine extra shoot days (asosplc.com, H1 FY26 interim results, 23 April 2026).

• Fully synthetic models carry the reputational cost. Mango, Guess and Gucci each ran one, and each drew public criticism.

• Disclosure did not prevent backlash in any documented case. Guess and Gucci both labelled their AI imagery and were criticised anyway.

• AI production is not automatically cheaper. The chief executive of the production company behind McDonald’s Netherlands’ pulled AI Christmas advert described weeks of work and “thousands of takes” shaped in the edit (thedrum.com, 11 December 2025).

• Two rules changed in 2026. Since 9 June 2026, New York requires a conspicuous disclosure in an ad with a synthetic performer. Article 50 of the EU AI Act has applied since 2 August 2026.

• The live legal risk is contract scope. One active US case turns on a retailer generating new images from a shoot whose contract permitted only minor edits.

What fashion brands using AI models shipped

The claims in this category are unusually unreliable, so the table below includes only cases with a primary or Tier-1 source, and separates what the brand did from what gets said about it.

Three names that circulate constantly do not belong on it. Adidas never ran the AI campaign attributed to it, which was a speculative piece by an independent filmmaker who said adidas never responded (ibm.com, 2025). We found no primary source showing Shein using AI-generated models in its own imagery. And we found no source in either direction on whether Levi’s ever deployed its pilot.

The split the benefit lists miss

Read the table by what sits behind the face and a pattern appears that the standard benefits listicle inverts.

The four volume retailers in the table did not put an invented person in their catalogs, on the record. Zalando and H&M work from replicas of real models. Zara edits photographs of models it has already shot, and two models told City AM they would be paid the same fee as a standard shoot. Whatever ASOS is doing internally, it described the output as model imagery produced alongside its own styling expertise rather than as synthetic people.

The three brands that put a fully synthetic human in market used it for a campaign, and none of them was solving a catalog problem. Gucci’s images promoted its Milan Fashion Week show. Guess bought an advertisement in Vogue. Mango made a campaign for its teen line.

So the form that vendors sell hardest, a person who does not exist, is the form the volume operators mostly did not buy. That is the single most useful fact in this category, and it follows from a straightforward constraint: at catalog scale you need the same face to be reliable, repeatable and defensible across thousands of SKUs and several years. A contracted human with a signed agreement is all three. A face regenerated from a prompt is none of them until you build the machinery to make it so.

What the volume operators bought

The clearest number in this category comes from a public company’s interim results rather than a vendor case study, which matters because filings carry a different standard of care.

ASOS reported that AI Studios uses generative AI alongside its styling expertise to create product imagery and content at scale, including model imagery, styled looks and video. For April’s campaign it “enabled the creation of more than 1,600 additional assets that would have required c.9 extra shoot days using traditional methods”, and it delivered 95% of SKUs for that campaign, which is why ASOS said it would roll the approach out more broadly across Own Brand products (asosplc.com, 23 April 2026).

Read what that sentence measures. It counts additional assets. It does not report a smaller invoice, so the gain ASOS published is volume.

Zalando’s disclosure has the same shape. Around 70% of its editorial campaign assets in Q4 2024 came from AI-driven content production, and its VP of content solutions, Matthias Haase, described moving from spotting a trend to producing tailored content in under 24 hours (corporate.zalando.com, 7 May 2025). Haase separately told Reuters the approach cut production time from six to eight weeks down to three or four days, with roughly a 90% cost reduction (reuters.com, May 2025). Note the attribution difference: the 70% is in Zalando’s own release, the 90% is an executive speaking to a reporter.

This is the same conclusion our own breakdown of what generative AI in fashion pays for reaches from the production side. The saving shows up as throughput, and throughput is what compounds.

One caution against assuming AI production is inherently cheap. When McDonald’s Netherlands pulled an AI-generated Christmas advert in December 2025, the chief executive of The Sweetshop, the production company behind it, described weeks of work and “thousands of takes” shaped in the edit “just as we would on any high-craft production” (thedrum.com, 11 December 2025). Reports at the time gave different figures for team size and weeks, so we do not repeat them. Generative tools move the labour rather than deleting it, and a campaign-grade result still costs campaign-grade effort.

Why fully synthetic is the expensive route

The subscription is the smallest line in this decision.

Every brand in the table that shipped a fully synthetic model drew public criticism. Mango’s teen campaign drew consumer complaints, including charges of false advertising (yahoo.com, July 2024). Guess and Vogue drew reader backlash after the advertisement, and we found no retraction by either. Vogue’s clarification was that the advertisement was paid, and that AI models had not appeared in its editorial pages (cnn.com, 31 July 2025). Gucci’s images drew public criticism as “AI slop” even though they were labelled (bbc.com, February 2026).

Worth stating plainly, because the opposite is repeated often: we found no brand in this set that withdrew AI-generated model imagery after backlash. Levi’s is the only formal correction, and it retracted the framing of its announcement rather than any image. Its editor’s note said the pilot was not “a means to advance diversity or as a substitute for the real action that must be taken”, and confirmed the company was “not scaling back our plans for live photo shoots, the use of live models, or our commitment to working with diverse models” (levistrauss.com, March 2023).

The legal exposure is also not where most articles put it. One active US case does not concern synthetic faces at all. A model sued a clothing retailer in New York on 22 May 2026, withdrew the case a week later and refiled it on 15 June, alleging the brand photographed her and then generated entirely new images in poses she never took. Her complaint states her contract permitted minor edits but “does not in any way authorize the creation of entirely new images, scenes, poses, or compositions” (petapixel.com and wwd.com, June 2026).

That is a contract-scope problem, and it is the one most likely to reach an ordinary brand. A release covering a shoot is not a release covering generation from that shoot. If you are editing or extending imagery of real people, the wording of your model release is the control that matters, and our guide to swapping the model in a product photo covers the consent question in more depth.

Labelling does not buy you cover

This is the finding that most changes how a brand should plan, and it is consistent across every case.

Guess disclosed, in fine print reading “Produced by Seraphinne Vallora on AI”. Gucci labelled its images “created with AI”. Both were criticised anyway, and the criticism was rarely about concealment. It was about the decision itself, about beauty standards, and about work moving away from photographers, stylists and models. Mango’s own announcement said nothing about labelling either way, and the claims we found that it labelled the campaign come from secondary sources.

Treat disclosure as a compliance obligation you have to meet, and plan for reputation separately. They are separate problems with separate solutions. A label addresses compliance. The second is only addressed by the choice of what you generate and who you still pay, which is exactly why the volume operators’ route reads better in public than the synthetic route does.

What you owe your audience as of September 2026

Two changes landed recently enough that most articles on this topic predate both.

This is general information, not legal advice. Article 50 of the EU AI Act has applied since 2 August 2026. Article 50(4) requires deployers to disclose deep fakes: realistic AI images, audio or video that could pass as real. Providers of the AI tools carry a separate machine-readable marking duty under Article 50(2) (European Commission FAQ, accessed September 2026). The Commission published guidelines on Article 50 on 20 July 2026, and they are not binding. The AI Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, did not move the Article 50 date. It gave tool makers until 2 December 2026 to add machine-readable marks to systems already on the market (European Commission, July 2026).

Separately, New York’s synthetic performer disclosure law took effect on 9 June 2026. It is General Business Law section 396-b. If you make an ad and you know it contains a synthetic performer, the ad must say so in a way people will notice. New York defines a synthetic performer as a digital asset, made with generative AI or software, that looks like a human performing, where the person is not recognisable as any real, identifiable performer. The civil penalty is $1,000 for a first violation and $5,000 for each later violation (nysenate.gov, accessed September 2026). The definition covers an “audiovisual and/or visual performance”. We found no published enforcement decision on how it applies to a still product photo.

Notice that both rules can reach an invented person. New York’s applies precisely because the person is invented. The Commission’s guidelines treat realistic AI-generated human avatars or personas as persons, and say it is enough that the person could plausibly exist. So a realistic synthetic model can count as a deep fake under the EU rule too. Usually the brand that uses the AI tool is the deployer. Do not treat “she is not a real person” as an exemption.

A third rule sits underneath both and does not depend on either. Since 19 June 2025, New York’s Fashion Workers Act has required separate written consent before a model’s digital replica is made or used. The consent must state the scope, purpose, pay and how long it will be used. Routine edits such as colour correction and minor retouching are not digital replicas (New York State Department of Labor, accessed September 2026).

Which specific assets in a drop need a label, and which do not, is its own question with its own answer. Our guide to labeling AI-generated fashion images works through it asset by asset.

The routes are not equivalent under these rules, and the differences are set out in full in our guide to where AI fashion models come from. Platform policy adds a third layer. TikTok requires creators to label AI-generated or heavily edited content that shows realistic people or scenes (TikTok Community Guidelines, accessed September 2026). In ads, TikTok asks for disclosure when images, video or audio are fully AI-generated or heavily AI-edited (TikTok Ads policy, accessed September 2026).

One honest counterweight. We found no regulator action against a brand for using AI-generated fashion models, as of September 2026. The UK advertising codes contain no AI-specific rules, and the ASA says its existing rules on misleading ads apply however an advert was made (ASA, accessed September 2026).

How to pick the form your brand needs

Work from the constraint rather than the tooling.

If the job is catalog volume, you want the pattern the volume operators used. Keep real, contracted people in the imagery, get consent that names AI generation with scope and duration, and put the gain into shot count per SKU rather than into a smaller production budget. This is the best-evidenced route and the least exposed.

If the job is a campaign moment, a fully synthetic model is viable, and three well-known brands have proven it ships. Budget for the reaction rather than assuming a label prevents it, and check the New York disclosure rule before the media plan is signed.

If the job is more angles, scenes and formats from shoots you already own, this is the cheapest route and the one with the live litigation. Read the release from the original shoot before generating anything new from it.

Before you commit to a route, put real numbers against your current baseline. Our breakdown of what a product photoshoot costs gives the comparison, the tool comparison covers what each category of vendor is built for, and if the open question is whether generated imagery converts as well as a booked shoot, that evidence is weighed in AI models against real models.

Catalog volume on DesignerBox

The brands above are named from their own disclosures and from press reports. None of them is presented here as a DesignerBox user, and none endorses it.

DesignerBox is AI creative production for agencies and brand teams, and it fits the first and third jobs above. Every result starts from your real product photo, so you check it against the garment you shot. The model is one step in the workflow, and you can change that step per shot. You see what a run costs before you start it. Basic is $15 a month billed monthly for 500 credits. The commercial licence starts on Pro, at $35 a month billed monthly, and virtual try-on starts on Premium, at $75 a month billed monthly. Pricing for every plan is on the pricing page.

The volume operators’ lesson applies to how you set this up. Build the shot once for one garment, with the model, the light, the framing and your brand rules fixed, and save it as a workflow. A saved workflow runs the same way on the next garment and the one after it, so nobody decides the look again. Three critic steps score the results, and best-of-N keeps the best one. Batch is coming: it will run one workflow over a whole sheet of products.

For teams producing a drop at a time, Dress my model puts a garment on a model, a model template lets you cast one face and reuse it, and DesignerBox for fashion brands shows the full production path. If you still work from real models, your release has to name AI generation before you use their photos.

Start from a template, add your brand and your garments, and run it. The cost is shown before the run.

FAQ

Which fashion brands use AI models?

ASOS, Zalando, H&M, Zara, Mango, Guess and Gucci all have documented programmes, but they differ fundamentally. Zalando, H&M and Zara built theirs on digital replicas or AI edits of real models, and ASOS has not said how its model imagery is made. Mango, Guess and Gucci published fully synthetic figures. We found no primary source for claims that Adidas, Nike or Shein use AI fashion models.

Do AI fashion models save money?

Sometimes, but the reported gain usually arrives as extra output rather than a smaller bill. ASOS counted more than 1,600 additional assets that would have needed around nine extra shoot days. It did not report a smaller bill. The production company behind one pulled AI advert described weeks of work and thousands of takes. Treat the saving as capacity, and do not plan on a discount.

Do I have to disclose that a model is AI-generated?

In the EU, likely, when the model is realistic. Article 50 of the AI Act has applied since 2 August 2026, and the Commission’s guidelines say an invented person can count as a deep fake when the person could plausibly exist. In New York, since 9 June 2026, an ad that you know contains a synthetic performer must say so in a way people will notice. The UK advertising codes have no AI-specific rule, and the ASA says its existing rules on misleading ads still apply. TikTok requires creators to label realistic AI-generated content. This is general information, not legal advice. Verify the current text before writing your own policy.

Is it legal to generate new images from an old photoshoot?

Only within the scope of your model release. One active New York case alleges a retailer generated entirely new images from a shoot whose contract permitted only minor edits. A release covering photography does not automatically cover generation. Get AI generation named explicitly, with scope and duration.

Are AI fashion models real people?

It depends on the route. A digital twin is a licensed replica of a real, contracted model. A fully synthetic model represents nobody. A third route uses a general-purpose image model with no provenance record at all. The route decides what consent and disclosure you owe.

Did any brand pull its AI model campaign after backlash?

We found no brand in the documented set that withdrew AI-generated model imagery. Levi’s issued a formal correction in 2023, but it retracted the diversity framing of its announcement rather than any image. We found no primary source for claims that other fashion brands pulled AI campaigns. The narrower production claim it kept, more body types shown more quickly, is the one the size-range research supports.

Does labelling protect a brand from criticism?

Not on the available evidence. Guess and Gucci both disclosed and were criticised anyway. Disclosure satisfies a legal obligation. It does not address objections to the decision itself, which is a separate problem solved by what you choose to generate and who you continue to pay.

Sources

• ASOS plc, interim results for the 26 weeks to 1 March 2026 (asosplc.com, 23 April 2026)
• Zalando, “Zalando explores digital twins: high-fidelity replicas of real models” (corporate.zalando.com, 7 May 2025)
• CNN, H&M plans to use AI clones of its models (cnn.com, 28 March 2025)
• The Drum, McDonald’s Netherlands pulls AI ad (thedrum.com, 11 December 2025)
• Reuters, Zalando AI content production (reuters.com, May 2025)
• Mango, first AI-generated campaign for its teen line (mangofashiongroup.com, July 2024), and consumer criticism coverage (yahoo.com, July 2024)
• Levi Strauss & Co., Lalaland.ai partnership post and editor’s note (levistrauss.com, March 2023)
• CNN and Forbes, Guess AI models in Vogue (cnn.com, 31 July 2025; forbes.com, 29 July 2025)
• BBC, Gucci AI-generated images (bbc.com, February 2026)
• City AM, Zara AI-edited model imagery (cityam.com, December 2025)
• PetaPixel and WWD, New York model lawsuit over AI-generated images, filed 22 May 2026 and refiled 15 June 2026 (petapixel.com, wwd.com, June 2026)
• European Commission, transparency obligations under Article 50 of the AI Act and AI Omnibus enters into force, accessed September 2026
• New York State Senate, General Business Law 396-b, synthetic performer disclosure, accessed September 2026
• New York State Department of Labor, Fashion Workers Act FAQs, accessed September 2026
• Advertising Standards Authority, disclosure of AI in advertising, accessed September 2026
• TikTok, Community Guidelines on AI-generated content and ads policy on misleading and false content, accessed September 2026
• DesignerBox pricing page (designerbox.ai/pricing), September 2026
• IBM Think, on the speculative adidas AI advertisement (ibm.com, 2025)

Brand programmes verified from company filings, corporate press releases and Tier-1 reporting as of August 2026, and re-checked September 2026. Regulatory positions verified from statute text and official guidance as of September 2026. This is general information, not legal advice. Verify current requirements before setting policy. Individual results vary.