top of page

Diverse Casting With AI Models: Representation, Backlash and Better Practice

3 days ago
9 min read

Updated: 12 hours ago

AI models can widen the range of bodies and skin tones shown in a fashion campaign, but they do not make a brand more diverse, and audiences have said so loudly. The documented backlash has the same shape each time: a brand presents synthetic people as a diversity measure, critics point out that real models of those backgrounds lose the booking, and the brand retreats to a narrower claim. The defensible use is also narrow. AI variants work for product views, where the question is how a garment looks across sizes or tones, and where real people were never going to be cast for every combination anyway. Real casting is needed wherever the image speaks for the brand: campaign leads, editorial, anything that signals who the brand is for. In every case the synthetic nature of the image should be disclosed, because the EU AI Act now requires it for realistic imagery.


Why this question comes up at all


Brands want to show their clothes on more kinds of people. Fit across sizes, colour on different skin tones, age ranges and body shapes are all things shoppers look for, and a traditional shoot covers few of them. Every additional model means another booking, another fitting, another day of styling, hair and make-up, and another round of approvals.


Generative production appears to remove that cost. One garment, many bodies, one afternoon. That is the pitch behind most searches for scalable AI solutions for diverse fashion campaigns, and it is not wrong as a description of what the technology can do. The difficulty is what the result communicates. A campaign image is not only a picture of a garment. It tells the viewer who the brand sees as its customer, and who it was willing to hire to say so.


Those two jobs, showing the product and signalling inclusion, are where the whole debate sits. Mixing them up is what has caused the backlash so far.


What has been documented


The cases below come from press coverage and brand statements as they were reported. They are not verdicts on the brands.


Levi's and Lalaland.ai (March 2023). The denim company announced a partnership with an AI model studio to test generated models showing a wider range of body types, sizes, ages and skin tones. Coverage at the time reported fast criticism: that the approach was a shortcut to the look of diversity, and that real models from those groups would miss out on work and pay. Within days Levi's issued a statement that the pilot was not a means to advance diversity and not a substitute for action on its diversity goals. It said it was not scaling back live shoots or its commitment to working with diverse human models, and described the technology as a way to publish more product images on a range of body types more quickly. NBC News reported that the episode also fed wider worries among fashion workers about being replaced, and that Levi's said the generated models would supplement, not replace, its shoots with live models.


Guess in Vogue (August 2025). Our article on how luxury brands use AI in campaign imagery covers the case in full. The part that matters here is that the criticism reported at the time combined three points: a small disclosure that many readers missed, unrealistic beauty standards, and a lack of diversity among the generated models. The synthetic models were not a diversity effort, yet they were read as a statement about who gets to represent fashion.


Mango (July 2024). The retailer published a campaign for a teen line that it described as fully AI-generated and said so openly. Trade coverage still raised ethical concerns about bias and unrealistic ideals in synthetic models. Disclosure answered one question and left the other open.


Three things follow from these cases. Generated people attract more criticism than generated backgrounds or products. Describing synthetic models as a diversity measure invites a response that real hiring is being displaced. And a label is not a defence: honest disclosure did not prevent criticism of the image itself.


Why synthetic diversity reads as a shortcut


The reaction is not irrational, and brand teams should understand its logic before deciding anything.


  • Someone is not hired. If a brand shows a plus-size model or a model of colour generated from a prompt, a real person of that description does not get the day rate, the credit or the portfolio image. The people most affected are often those who already find fewer bookings.

  • The image is not a person. Representation matters to viewers because someone like them was chosen. A generated figure is chosen by nobody. It can look inclusive without any inclusion having happened in the business.

  • It is cheap. Critics notice when the stated value, inclusion, coincides with the lowest cost route. Consultation with the people shown, or a plan to hire them elsewhere in the business, changes how that is read.

  • Generated faces repeat patterns. Systems trained on skewed material reproduce skewed ideals, including narrow beauty standards inside each group they are asked to show. This is a quality issue before it is an ethical one, and it is visible in review.


Synthetic imagery is not off limits. The claim a brand makes about it just has to be small enough to be true.


Where AI variants are defensible


The legitimate use is about the product, not the pitch to the audience. A shopper deciding on a size wants to see how a garment falls on a body like theirs. A shoot can realistically cast a handful of people. Generative production can show the same garment across a wider range, as long as the garment itself is rendered faithfully.


In practice that means:


  • Product pages and size guides, where the aim is fit information and the image is understood as a visualisation.

  • Colourways and tone checks for items where the colour reads differently against different skin tones, shown as additional views next to real campaign imagery.

  • Internal and early-stage work: line-sheet and pre-order visuals before samples exist. The same logic applies in B2B settings, as described in our piece on AI imagery for wholesale.

  • Market adaptation, where an existing campaign needs variants for markets the shoot could not cover, with approval from the people responsible for those markets.


Two conditions apply to all of these. The garment must look as it would in reality: fabric, drape and fit are product claims, and exaggerating them can mislead customers. And the viewer should not be led to believe that a real person of that description was cast.


Where real casting is needed


Some images carry the brand's identity, and for those a generated figure undermines the very thing being shown.


  • Campaign leads and hero imagery. The people in the main campaign are the brand's public statement about who it is for.

  • Editorial and press work, where credibility with the media depends on real people and real credits.

  • Anything described as inclusion. If a campaign is presented as celebrating a group, the people shown should be members of it, hired and paid.

  • Imagery of identity. Cultural dress, religious dress, and features tied to a community call for the involvement of that community, not an approximation.


A simple internal test helps: would the brand be comfortable putting the sentence "this campaign features generated people" at the top of the page and next to the credits? If not, the image belongs in a real shoot.


Disclosure, and what the law requires in the EU


Disclosure is the minimum, and it has a legal side. Since 2 August 2026, Article 50 of the EU AI Act requires deployers of AI systems to disclose content that constitutes a deep fake, meaning generated or manipulated imagery that resembles real persons and could be mistaken for authentic. Photorealistic campaign imagery of synthetic people can fall under that definition. We explain the rule and its limits in our guide to the EU AI Act labelling rules for ad imagery. This is general information and not legal advice. Whether a particular campaign is covered is a question for counsel.


Beyond the legal minimum, disclosure works better when it is specific. "Created with AI" tells a viewer little. A line saying that the models in this size range are generated, and why, tells them what they are looking at. The Guess case shows what happens when a label is small enough to miss: the reaction is then about the hiding as much as the image.


A practical approach for brands and agencies


1. Decide the job before choosing the method. Is the image meant to show fit, or to say something about the brand? Fit can be visualised. Statements need people.


2. Keep real casting in the plan. State in the brief that generated variants supplement live models and do not replace them, and keep the live casting budget visible. The Levi's statement shows how quickly this becomes the central question.


3. Do not call it a diversity strategy. Describe it as extended product visualisation. If diversity is a goal, it belongs to hiring, casting and product range, where it can be shown.


4. Review generated people for sameness. Look for the same face, proportions and styling repeated across the range. It is a common failure and a cheap one to catch.


5. Check the garment, not only the person. Fit and fabric accuracy per size are the claims the shopper relies on. Compare the output with the physical sample.


6. Agree rights early. If a real model's likeness is used as a base, the terms need to be in writing. See model likeness rights in AI campaigns and how digital twins of real models are made, licensed and paid.


7. Disclose clearly and specifically, in the place where the viewer sees the image, and keep an internal record of which images are generated.


8. Ask the people concerned. Where a campaign speaks to a community, involve members of it in review, with pay, before publication.


What this means for production partners


A production studio can build large numbers of variants. Whether it should is a decision for the brand and its agency, and a good partner will raise the questions above before it starts. The useful contribution is control of the product rendering, consistent quality across the range and a clean record of what is generated, so that the brand can answer questions after launch. The decision about who appears in the campaign's lead roles stays where it belongs, with the people who hire.


Frequently Asked Questions


Can AI models make a fashion campaign more diverse?

They can widen the range of bodies and skin tones shown, but they cannot make a brand more diverse. Diversity is about who is hired, cast and served, and a generated figure involves none of that. Brands that presented synthetic models as a diversity measure, such as Levi's in 2023, drew criticism and clarified that the models were not a substitute for real action.


Is it acceptable to use AI models for size and skin-tone variants?

It can be, for product visualisation where the goal is to show fit or colour across more people than a shoot could cast. The garment has to be rendered faithfully, the variants should sit alongside real campaign imagery rather than replace it, and the viewer should not be led to think a real person was cast.


When should a brand use real models instead?

For campaign leads, editorial work, and any image that signals who the brand is for. If a campaign is framed as celebrating a group, its members should be hired and paid. A useful test is whether the brand would be comfortable stating next to the credits that the people shown are generated.


Do AI-generated models have to be disclosed?

In the EU, Article 50 of the AI Act has required deployers to disclose deep fake content since 2 August 2026, and photorealistic imagery of synthetic people can meet that definition. Whether a given campaign is covered is a question for counsel. Beyond the law, clear and specific disclosure reduces the chance that the reaction is about concealment.


Does disclosing AI use prevent backlash?

No. Several brands that labelled their AI use were criticised anyway, because the objection was to the image or to the loss of work, not only to the lack of a label. Disclosure is a minimum, and it works best alongside a claim that is modest and true.


How do I check that generated models do not look the same?

Review the whole range side by side. Look for repeated faces, proportions, poses and styling, and for narrow beauty ideals within each group shown. Compare the garment on every variant with the physical sample. Both checks are inexpensive and catch most problems before publication.


Planning a campaign with AI imagery or video? Request an AI campaign.


About The Bridges AI

The Bridges AI is a hybrid AI creative agency: a creative studio and production partner for fashion, beauty, luxury and hospitality brands and for the agencies that work with them. It combines booked real talent, real products and human creative direction with AI production, and offers licensed digital twins of real models, only of models under contract. The Bridges AI works in strategic partnership with Most Wanted Models®, the German model agency founded in 1997, which brings nearly 30 years of experience in bookings, buyouts and usage rights.


Written by the editorial team of The Bridges AI. Last updated November 2026.


Comments


bottom of page