Is fashion AI biased? Sizes, skin tones and fairness explained

KEY TAKEAWAYS Summary by the editors

1. Bias in fashion AI arises when training data, design choices or evaluation under-represent certain body sizes, skin tones, ages or genders, leading to worse results for those customers.
2. The 2018 Gender Shades study found commercial facial analysis systems performed worst on darker-skinned women, with accuracy as low as 65.3% compared with 99.7% for lighter-skinned men in one system.
3. A 2025 study presented at NAACL analysed 4,000 images generated by DALL-E 3 to examine anti-fat and pro-thin bias in AI-generated images.
4. Tools such as Google's 10-shade Monk Skin Tone Scale help teams test whether image systems work across skin tones.
5. The Digital Omnibus on AI added Article 4a to the EU AI Act, allowing processing of sensitive personal data where strictly necessary to detect and correct bias, under safeguards.

Fashion AI can be biased, and the risk is highest where nobody tests for it. When training data and design choices centre narrow body sizes, lighter skin tones or particular ages, AI-generated images, virtual try-on, size recommendations and search can work worse for everyone else. Bias is not inevitable, but reducing it takes deliberate data collection, testing and human review.

Where does bias in fashion AI come from?

• Training data: fashion imagery has historically featured a narrow range of bodies and skin tones, and models learn those patterns.
• Product data: if extended sizes are photographed less often or described less fully, AI has less to learn from.
• Customer data: recommendation and fit models learn from past purchases and returns, which reflect what was offered and to whom.
• Design choices: the range of body types offered in a virtual model tool or avatar builder sets limits from the start.
• Evaluation: if a system is tested mainly on average cases, failures for other groups go unnoticed.

What does research show about bias in image AI?

The best-known evidence comes from facial analysis. The Gender Shades study by Joy Buolamwini and Timnit Gebru, published in 2018, evaluated gender classification systems from IBM, Microsoft and Face++ on 1,270 images. All three performed better on lighter-skinned than darker-skinned subjects, and worst on darker-skinned women. In IBM's system, accuracy was 99.7% for lighter-skinned men and 65.3% for darker-skinned women. The authors' conclusion was that inclusion requires intention.

Generative AI raises similar questions for body size. A 2025 paper presented in the Findings of NAACL, Decoding Fatphobia, examined anti-fat and pro-thin bias in images generated by DALL-E 3, using 4,000 images created from contrasting prompts and coded manually for body weight. The authors note that AI-generated images tend to reinforce social biases. For fashion, where generated models and campaign images are increasingly common, such patterns matter commercially as well as ethically.

Read also

EU AI Act: what do fashion companies need to know in 2026?

Which fashion AI applications are most exposed to bias?

Why does size bias matter commercially?

Bias is not only an ethical issue. When a size recommendation is less accurate for some customers, they are more likely to receive items that do not fit, which drives returns and lost sales. When AI-generated product imagery shows garments only on one body type, customers of other sizes have less information to judge fit and drape. When search or recommendation systems learn from past sales, ranges that were under-stocked or poorly presented can be shown even less, reinforcing the problem.

These effects are measurable. Comparing return rates, conversion and customer complaints across size groups and regions can reveal where AI tools perform unevenly. Such analysis should be part of every evaluation, alongside overall accuracy figures supplied by vendors.

Can AI make fashion more inclusive?

AI can also be used to show more diversity. In April 2023 Levi's announced it would test AI-generated models with different body types, ages and skin tones together with the Dutch digital fashion firm Lalaland.ai. After criticism that the plan amounted to artificial diversity, Levi's clarified that the pilot was an experiment and that AI would likely never fully replace human models for the company. The episode shows that inclusion through AI can be perceived as a shortcut if it replaces real representation rather than supporting it.

Measurement tools help. In May 2022 Google released the Monk Skin Tone Scale, a 10-shade scale developed with Harvard sociologist Ellis Monk, and made it openly available for research and product development. Google uses it to evaluate computer vision models for fairness across skin tones. Fashion teams can use similar scales to structure test sets for imagery, try-on and visual search.

What does EU regulation say about AI bias?

The EU AI Act addresses bias most directly for high-risk systems, such as those used in recruitment, and through data governance requirements. The Digital Omnibus on AI, Regulation (EU) 2026/1744, added Article 4a, which allows providers and deployers to process special categories of personal data where strictly necessary to ensure bias detection and correction, subject to safeguards such as pseudonymisation, restricted re-use and strict access controls. That matters because testing for bias by skin tone or other sensitive characteristics can involve exactly such data.

Most fashion AI applications, such as imagery and recommendations, are not high-risk under the AI Act. But discrimination, consumer protection and data protection law still apply, and biased outputs can cause reputational damage quickly.

Read also

Who owns an AI-generated fashion design? Copyright and IP explained

How can fashion brands reduce bias in AI?

1. Define which groups matter for each application, such as size ranges, skin tones, ages and genders.
2. Build representative test sets and measure performance for each group, not only on average.
3. Ask vendors how their models were trained and tested for bias, and request results.
4. Involve diverse reviewers and, where possible, customers in evaluating outputs.
5. Monitor results after launch, for example returns by size group, and retest after updates.
6. Use AI imagery to complement, not replace, real representation in campaigns.

Frequently asked questions

Why are AI fashion models often thin and light-skinned?

Generative models learn from existing images, and fashion imagery has historically shown a narrow range of bodies and skin tones. Without deliberate prompting, data choices and review, outputs tend to reproduce those patterns.

How can you test fashion AI for skin tone bias?

Build a test set covering a range of skin tones, for example structured with the 10-shade Monk Skin Tone Scale, and compare output quality and error rates across groups. Repeat the test after every model update.

Do size recommendation tools work for all sizes?

They can be less accurate for sizes with less purchase and return data, often the smallest and largest. Brands should measure accuracy and return rates by size group and improve data where results are weaker.

Does the EU AI Act regulate bias in fashion AI?

Bias rules are strictest for high-risk systems such as recruitment AI. Since the Digital Omnibus, Article 4a allows sensitive data to be processed where strictly necessary to detect and correct bias. Most fashion imagery and recommendation tools are not high-risk, but other EU laws on discrimination and consumer protection still apply.

GuideThe complete guide to AI strategy for fashion companiesRead the complete guide