Inclusivity and epistemology in AI
The recent news of the Garrick Club – one of the oldest and most exclusive private clubs in London — admitting women for the first time in its nearly 200-year history was rather surprising. It’s surprising to find that in 2024 there are institutions that openly deny access to women. The decision is, of course, positive for inclusivity and equality: its doors are now open to talented women to access those networks and resources that were previously unavailable. But how will the Club integrate women into an established culture where a sense of male privilege has dominated from its conception? From new restroom facilities to changes on the bar stock, surely opening the membership to women and making them feel welcome must cause a significant disruption to the Club’s management team.
Most importantly, who would want to join the Garrick Club now as it transitions to becoming a more representative and dynamic space; a place that will probably maintain its invisible legacy and status quo for years to come but with a few women sprinkled in? Wait a minute – did I just describe the Technology/IT sector?
The Garrick Club offers a good simile to illustrate the male-dominated atmosphere often found in academia and industry, in particularly AI, ML and Data Science fields. When knowledge is built from unequal relationships, hidden harms are caused to those that do not fit in. Those that do not feel welcome – the minority, the underrepresented, the token – are not valued or included to the same extent. No wonder women working in AI and data science in the tech sector have higher turnover and attrition rates than men.
Including women and other underrepresented groups in the field of AI is about more than just increasing the count of diverse bodies. What we have to recognise is that previously excluded groups have knowledge, skills and crucial lived experience that can help to solve some of AI’s most pressing challenges. Yet the deliberate exclusion of certain subjects, methods and knowledge – by classing them as ‘too social’, ‘not technical’, or ‘too subjective’ – will inevitably result in an impoverishment of the epistemic system, which loses out on multi-disciplinary advancement and the capacity for self-criticism. To date, there is a striking scarcity of diverse high-quality datasets essential to interrogate and tackle inequalities in AI. This is not surprising; this is expected and consequential from an environment that resembles the Garrick Club and its brand-new inclusion drive. High quality data is an important aspect, but equity in AI is about so much more than datasets. It’s about how those datasets are constructed, the responsible frameworks contextualising that data, the benchmarks used to evaluate the quality of AI-based decisions, the questions we ask of data, and the limitations of our methods in answering them. It is just as much about the blindspots that existing, well-accepted, AI methods produce, as it is about the methods themselves. Only with a committed community of researchers, aware of the diversity issues currently impacting AI, blindspots will be addressed.
There is a point when cumulative evidence reaches an impasse or saturation point. There is no need to repeat the same message. There is little new to learn when reading reports from the House of Lords Select Committees on AI ‘advocating for an increase in gender and ethnic diversity amongst AI developers’, or European Commissions noting that it is ‘high time to reflect specifically on the interplay between AI and gender equity’. This is a fundamental ethical issue of social and economic justice, as well as one of value-in-diversity and responsibility. The lack of gender diversity in AI negatively affects the production of knowledge and prevents women from participating in meaning-making and meaning-sharing practices, influencing strategic decisions and sharing power. We need to value multi-disciplinary contributions as highly as the deeply technical, and we can no longer view the deeply technical as sufficient for progressing the field of AI.
Gender diversity and methodological variety could be viewed as two sides of the same coin. By combining a diverse and methodologically rich environment for AI, we can then create AI that is fair, effective, responsible, and one that tackles society’s challenges from multiple angles, ensuring a plurality of benefits.
The newly awarded RAI UK Keystone projects are all lead by women. Gender was not a factor in their selection. RAI UK’s diverse leadership, its commitment to EDI values as a selection criterion to assess proposals, and a representative Equities Committee contributed to the fact that qualified and ambitious women applied for these substantial awards. Supportive processes are an effective and proactive way to tackle gender gaps and equity in AI, and the inclusion of women in leading roles can influence the design and development of responsible AI, gender-sensitive design and implementation of data science and ML techniques. In time, this can lead to deliverables that are more inclusive, representative and pluralistic. From an equity viewpoint, this is a step forward.
Elvira Perez Vallejos (Chair Ethics Pillar)
Kate Devlin (Deputy Chair Public Engagement, Outreach and Policy Pillar)
Caitlin Bentley (Deputy Chair Skills Pillar)