Artificial Intelligence (AI) has never been more top of mind – and for good reason. Industry experts are both excited and inspired by the dizzying possibilities it poses, but most are also cautious about the impact of the wide-scale use of this technology, including how it will extend toward gender bias in generative AI. It has become accessible and democratised to anyone with an internet connection, and many are calling for regulation, education around the ethics of engaging with it, and for legislative considerations around intellectual property. For many women (and women in AI) however, AI models are an extension of the unconscious gender biases that have existed since the early days of the internet…with generative AI as a powerful visual remnant of these ingrained gender biases.

Table of contents

• What is generative AI?
• How does gender bias manifest in generative AI?
• Addressing bias in generative AI
• Conclusion

What is generative AI?

Generative AI is defined as a type of artificial intelligence system capable of generating text, images, or other media in response to prompts; two popular examples are OpenAI Dall-E, and the text-to-image feature on Canva, the renowned SaaS-powered design platform. A user can, for example, type in a description and select a visual style, and the AI tool will generate an image accordingly.

The Deloitte article ‘Generative AI is all the rage’, explains that the main difference between “traditional AI” and Generative AI is that in the latter, the output is of a higher complexity. Rather than just a number or a label, the output can be an entire high-resolution image, a full page of newly written text (which is generated word by word), or any other digital artifact. “This introduces an interesting new element: There is usually more than one possible correct answer. This results in a large degree of freedom and variability, which can be interpreted as creativity.”

Gender bias in AI-generated content

 The nature of this technology promises vast implications for creative industries like graphic design, photography, fine arts and copywriting, but may ultimately lead to a regression in the representation of women in the media. According to the Centre for International Governance Innovation, most text-to-image models are trained on LAION-5B, a large open-source data set compiled by scraping content, including images, from the internet. But the internet lacks gender-representative data sets and is littered with mis- and disinformation – which ultimately means generative AI tools are being trained on and shaped by flawed data.

Alex Dowdalls, Managing Director at AXVECO, points out that many generative AI tools have also been trained with imagery of women who look like models with perfect features, which reinforces unrealistic societal expectations about their looks in relation to their value. In a recent workshop, he demonstrated the use of Chat GPT prompts to generate an image in Midjourney, a popular tool amongst marketing communication agencies. In his prompt, he requested a French baker with the following description: "Every wrinkle and crease on his face tells a story of years spent perfecting his craft, while the intensity of his gaze conveys a passion for baking that cannot be contained ... drawing attention to the baker's expressive eyes and the delicate lines etched on his weathered face." When replacing the male pronouns with female pronouns, he got a vastly different result:

Alex labels bias as the ‘Achilles heel of AI’: “If the data we use remains biased, then we will never get unbiased outcomes.” His passion for this subject is why he supports and lectures at programmes that educate about the challenges of AI as well as the perceived benefits to aspiring female digital leaders, which is one of the outcomes in the module ‘A New Age of Competition’ of the RightBrains Digital Leadership Programme.

Addressing gender bias in generative AI

The creation of diverse teams within AI and tech companies is a logical starting point when considering ways to address gender bias in generative AI, but it’s an industry that is still male-dominated. Organisations like RightBrains, with the mission to help bridge the gender gap in digital technology, hope to make a difference on this front by bringing and retaining more women within the digital technology industry. Hilary Richters, lead of the Deloitte Digital Ethics team and another lecturer of this year’s RightBrains Digital Leadership Programme, adds: “To create a more digitally ethical industry (and society), we need every possible skill, perspective, and vision. Women, who have been historically underrepresented in this sector, can bring fresh perspectives to the table. Ultimately, it will be beneficial to cover as many potential blind spots as possible through diverse leadership in digital technology.”

Interestingly, when prompted to advise on this issue, Chat GPT itself mentions that it’s also critical that companies should collect diverse, representative and ethical data during the training phase of AI models, together with focused algorithmic improvements.

It’s good to know that there are also initiatives underway like MissJourney, an AI alternative that creates artwork exclusively of women, with the aim of actively countering current biased image generators and promoting inclusive digital realities. MissJourney marked the start of the year-long TEDxAmsterdam Women theme ‘Decoding the Future’ -- and it’s imperative that initiatives like these should be supported and promoted early in the AI boom.

Conclusion

Gender bias in generative AI is an issue that society will need to navigate to avoid regression in terms of representation of women in the media – and ultimately within the industry. Although diverse teams are a solid way to mitigate the potentially damaging effects of gender bias in generative AI, there are challenges in terms of gender balance within the digital technology industry. Through education, information and better data sets, generative AI can become an even more powerful tool propelling industries forward.