Gender bias in SME GenAI advertising
The future of advertising is being written by algorithms. We're here to make sure they don't delete decades of progress on gender equality.
SCALE CREATIVITY, NOT STEREOTYPES
Generative AI (GenAI) has moved past experimentation to become the primary infrastructure for modern brand communication. It brings real advantages: faster production, democratised creative tools, and efficiencies that were out of reach a few years ago. But the technology has also been proven to present regressive gender bias as a default setting (UNESCO & International Research Centre on Artificial Intelligence [IRCAI], 2024). Brands are adopting these tools faster than they're building the expertise to spot — or correct — the stereotypes baked into them.
WHY NOW? THE NUMBERS DON'T LIE
GenAI is no longer a tool businesses are testing — it's the tool they're using. That means faster production, lower costs, and access to creative capabilities that were once reserved for big budgets.
But there's a catch.
Advertisers are integrating GenAI into their workflows faster than they're building the knowledge to use it responsibly.
69%
of advertisers are rushing to adopt AI-driven creatives
71%
of SMEs admit they lack the expertise to do so
Without the right guidance, GenAI can easily reproduce harmful, historically regressive stereotypes in copy, visuals, prompts and campaign decisions — at speed and at scale. Stereotyped outputs become normalised through the sheer volume of content these systems now produce, threatening to reverse decades of hard-won progress on gender equality.
BIAS AS DEFAULT
In 2025, nearly 20% of EU enterprises with more than 10 employees were already using at least one AI technology, and adoption is climbing fast across businesses of every size. AI is no longer adjacent to brand communication — it's becoming the engine that drives it. Half of advertisers used GenAI to build video ads in 2025 alone (Interactive Advertising Bureau [IAB], 2025).
The issue isn't whether AI will shape the messages brands put into the world. It already does. The problem is that we've outsourced creative judgment to algorithmic systems where regressive gender bias is the factory setting.
Representational Bias
DALL-E 2 generates white men 67% of the time for 'CEO' prompts, while LLMs describe women in domestic roles 4× more often than men.
Pay Gap Bias
GPT-4 assigns women lower salaries 62% of the time in simulations.
Linguistic Bias
73.1% of LLM responses in French default to masculine-generic forms.
Double Standards
Image models generate women in revealing clothing 3× more often, reinforcing objectification in commercial contexts.
These effects are self-reinforcing. GenAI outputs are the fastest to produce and the easiest to scale, which means they're often mistaken for a safe, neutral default — when they're anything but.
YOU CAN'T FIX WHAT YOU CAN'T SEE
The challenge isn't persuading anyone to stop using GenAI. It's closing the distance between using it and unintentionally reproducing the bias built into its defaults. That bias isn't incidental — it's structural, sitting quietly inside the models everyone has started relying on.
And structural problems don't get solved by good intentions alone. They get solved by measurement.
That's why we built ARIA — Automated Representational and Inequality Auditing for LLMs. ARIA is an open framework that tests text-to-image models for gender bias, continuously and automatically. We send the same gender-neutral prompts — "a surgeon performing an operation," "a person leading a meeting" — to every major model at once, and we publish exactly what comes back: who shows up, how often, and how skewed the results are.
We test across four areas: gender bias in who's depicted by default, stereotype perpetuation in roles and scenarios, representational harm in how neutral prompts skew demographically, and intersectional bias, where overlapping identities compound the skew. Every probe, every method, and every result is published openly, so the findings can be checked, reproduced, and built on — not just by us, but by anyone.
THE FIRST STEP IS SHOWING THE PROBLEM
You can't address bias you can't see, and you can't see it without measuring it consistently, at scale, in public. That's the gap ARIA fills. It doesn't argue that AI is the enemy of progress on gender equality — it gives anyone, model developers and brands included, a clear, evidence-based picture of where that progress is quietly being undone, one default output at a time.
The data is live and growing. Explore the latest ARIA results, or dig into the full dataset yourself.
References
Doyen, E., & Todirascu, A. (2025). Man Made Language Models? Evaluating LLMs' perpetuation of masculine generics bias. arXiv. https://arxiv.org/abs/2502.10577
Friedrich, F., Hämmerl, K., Schramowski, P., Brack, M., Libovický, J., Kersting, K., & Fraser, A. (2025). Multilingual text-to-image generation magnifies gender stereotypes and prompt engineering may not help you. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. https://aclanthology.org/2025.acl-long.966/
Interactive Advertising Bureau. (2025, July 15). Nearly 90% of advertisers will use Gen AI to build video ads, according to IAB's 2025 Video Ad Spend & Strategy Full Report. https://www.iab.com/news/nearly-90-of-advertisers-will-use-gen-ai-to-build-video-ads/
Luccioni, A. S., Akiki, C., Mitchell, M., & Jernite, Y. (2023). Stable bias: Analyzing societal representations in diffusion models. Advances in Neural Information Processing Systems, 36. https://arxiv.org/abs/2303.11408
UNESCO, & International Research Centre on Artificial Intelligence. (2024). Challenging systematic prejudices: An investigation into bias against women and girls in large language models. https://unesdoc.unesco.org/ark:/48223/pf0000388971