Dr. Rebecca Swift, Senior Vice President of Creative at Getty Images discusses how companies should be conscious of the imagery they use in communications.
A growing area of discussion is how AI contributes – or negatively impacts – gender equality. If AI is to progress how women are visualised, we must first confront the ‘sins of the past’ embedded in its training data – and then make deliberate choices to break that cycle.
Mainstream generative AI systems are trained on decades of imagery scraped from the open web. That archive – shaped by commercial choices, cultural biases, and narrow ideas about who gets seen and how – has defined the visual shorthand for ‘woman’ for more than 30 years. When models absorb that history wholesale, they don’t just reflect it; they reproduce it at scale. In other words, AI gives historic bias new speed and reach, unless we intervene with intention.
AI learns brands
This matters because what AI learns can become what brands, publishers, and institutions will publish next if using these tools. If the past said ‘woman at work’ means a young, white assistant in a supporting role, AI will continue to produce that image – again and again. And if the web has favoured younger women over older women, or framed ethnicity as a token rather than a dimension of identity, those patterns risk being amplified in AI generated outputs. Without intervention, the past becomes the future on repeat.
There is, however, real progress to build on. Over the last eight years, commercial storytelling has improved markedly, powered by ever more authentic and high-quality pre-shot (stock) imagery available to it. Today, businesses have access to visuals that authentically reflect the breadth of women’s lived experience across gender identity, age, ethnicity, culture, disability, body shape, profession, and context. We increasingly see older women in positions of authority, women with disabilities in imagery that focuses on what they are doing rather than what disability they have and ethnicity featured as a background dimension of a woman’s identity rather than token call out. These are meaningful shifts.
Yet our own VisualGPS research shows that this more inclusive supply of images does not consistently translate into the images people choose to use in company communications. In Financial Services, visuals popular with EMEA businesses still centre on a single archetype: the young, white, millennial woman. Women over 60 appear in just 2% of visuals downloaded by financial services businesses – despite being among the most powerful financial decision-makers across Europe – and are typically not shown across broader financial realities where they save, invest, plan, and lead.
Generation X women are similarly underrepresented. VisualGPS data shows that just 7% of European Gen X women are depicted as business owners in Financial Services visuals popular with European brands over the past year. Representation of leadership, entrepreneurship, hybrid working, frontline roles, and decision-making moments remains far too limited – out of step with reality and audience expectations.
If the inputs are dominated by the sins of the past, the outputs will echo them.”
This is not confined to a single category; there are specific issues by category. In healthcare imagery, women in their twenties appear twice as often as women aged 40-59. In automotive, there is a lack of women from a diversity of ethnic backgrounds or women as engineers or mechanics. The issue is clear: despite wide availability of crafted, authentic visuals, usage choices often fall back on familiar archetypes. Those choices then circulate in marketing online, where they are scraped, learned, and reproduced – risking yet another turn of the wheel.
The impact of AI ‘slop’
Compounding this challenge is the proliferation of low quality, low value generative content – what many call AI ‘slop’. When models are trained indiscriminately on scraped imagery, research shows that output skews toward clichés and overrepresented tropes. In practice, that means reinforcing outdated stereotypes of women and limited or token visibility of ethnicity, age, or disability. Put simply: if the inputs are dominated by the sins of the past, the outputs will echo them.
But this is also where the opportunity lies. Input determines output. With the right tool and the right practice, AI can help correct historic visual bias rather than compound it. That begins with how models are trained and what they are trained on. Tools built from licensed, carefully curated datasets start from a stronger foundation. Our own generative tool – trained from content within our creative collections designed to broaden representation across communities and under-represented groups, is one such example. Its inputs and therefore outputs are representative by design.
Intentional use matters just as much as responsible training. Prompts should be explicit about age, ethnicity, body type, disability, role, context, and agency. Guardrails – like inclusive taxonomies, prompt guidance and an inclusive human review panel – help avoid stereotypical outcomes. And equity-minded workflows ensure the images chosen for campaigns and communications reflect the breadth of audiences served.
Making better choices
Building equity in visual storytelling requires making better choices. Sometimes that means turning to responsibly trained AI. But often, the most effective step is choosing real, current, human‑crafted imagery that authentically captures women’s diversity.
AI cannot – and should not – replace the richness of human‑created photography. Instead, it should complement it, helping us broaden representation rather than flatten it. With intention on both fronts, we can move beyond the sins of the past and build a visual future worthy of the women we seek to represent.
