As we approach Women’s History Month in March, it seems pertinent to address the key challenge related to gender bias within AI, particularly large language models (LLMs), that produce outputs that reflect or reinforce stereotypes, inequalities, or unfair assumptions about gender. For example: associating certain professions such as nurse or data scientist with specific genders, using language that reinforces roles such as “he” for lawyer and “she” for teacher or generating harmful or stereotypical content about gender identities such as “women are bad drivers.”
Large Language Models (LLMs) such as ChatGPT, Llama 2 etc. have inherent gender bias within these models as per the UNESCO report (2024) on “Challenging Systematic Prejudices: An Investigation into Gender Bias in Large Language Models.” These LLMs may associate certain traits, roles, or behaviours with specific genders leading to them suggesting that women are more suited for caregiving roles while men are more suited for leadership roles. These models might default to using male pronouns for generic roles or use gendered terms inappropriately. These models may also struggle to handle non-binary or gender-neutral language and may default to binary gender constructs. If the training data for these models contains sexist, misogynistic or discriminatory content, then these models may reproduce or amplify it.
Gender bias in these models is caused by factors including training data with large datasets scraped from the internet leading to biased or stereotypical content. This training data may also include lack of representation of certain genders or identities that may be underrepresented in the training data leading to skewed outputs. The way in which these models are designed and fine-tuned may also inadvertently reinforce biases. Furthermore, if human reviewers introduce biases during fine-tuning these models, then these biases can be incorporated permanently into these models.
There are massive implications for wider society and the world at large for this gender bias in these AI models leading to gender discrimination in recruitment and hiring practices, biases in determining credit scoring and loan approvals, job displacement in industries where women form a large part of the workforce, reinforcement of harmful stereotypes and societal inequalities, exclusion, discrimination and underrepresentation of women and marginalised groups, erosion of trust is AI systems if they are perceived as biased or unfair amongst other areas.
UNESCO Member States adopted the recommendations made by UNESCO on the ethics of AI detailed in the United Nations report (2024) on “Governing AI for Humanity” on the use of responsible and ethical AI. These recommendations as well as my additional proposed solutions below are aimed at mitigating the inherent gender bias in AI models:
- Improving training data by ensuring diverse representation of genders and identities in the training data and removing harmful or biased content,
- Developing tools to detect and measure bias in model outputs and using techniques to reduce gender bias,
- Involving human oversight in the development and evaluation of LLMs, identifying and correcting biased outputs during fine-tuning,
- Documenting biases and limitations in model documentation,
- Encouraging open research and collaboration to address bias,
- Allowing users to specify preferences for gender-neutral or inclusive language,
- Funding gender parity schemes in companies,
- Financially incentivising women’s entrepreneurship (since only 2% of women founded businesses receive venture capital funding as per google for startups for female founders),
- Investing in targeted programmes to increase the participation of women in STEM (science, technology, engineering, and mathematics) and,
- Diversifying recruitment in companies at all levels.
- Implementing responsible AI regulation and wider regulatory framework that incorporates mitigating gender bias in AI.
These recommendations are crucial as gender bias in AI models can perpetuate harmful stereotypes and inequalities. It can be challenging to address gender bias in AI models due to barriers that are deeply ingrained within language and culture making it difficult to fully eliminate. Also, efforts to reduce this bias can sometimes lead to overcorrection or unintended consequences and societal understanding of gender is constantly evolving and AI models must adapt to these changes.
By undertaking a holistic approach to addressing gender bias in AI models at multiple levels: data, model development, user interaction, responsible AI regulation, organisational policies, advocacy, and systemic societal issues. We can build fairer and more inclusive AI systems by not only improving AI fairness but also contributing to broader societal progress towards gender equality.
Moving forward addressing gender bias in AI models must be undertaken at a global level by reaching consensus on a responsible, fair, balanced, transparent, innovation-friendly, principles-based, and coherent regulatory framework. As well as prioritising diversity in AI development teams, engaging with affected communities to understand their needs and concerns, continuously monitoring and updating models to reflect evolving societal norms and implementation of risk mitigation procedures to address gender bias in AI models. So that we can create a global society where AI works for the greater public good by being more equitable, beneficial, and inclusive for all members of society.
