Why inclusive AI matters | Chenayi Mutambasere | TEDxSt Helier

Key Concepts:

• Artificial Intelligence (AI) bias
• Data inclusion/exclusion
• Representation gap in AI
• Language gap in AI
• Digital exclusion/hunger/poverty
• Data as an economic asset
• Inclusive AI design
• Global data standards and policies

1. AI and Data Dependency:

• AI is likened to a "smart intelligent mirror" that reflects the data it's fed.
• AI's effectiveness is limited by the data it receives; it's not "magic."
• Unseen faces and unheard voices are not reflected in the AI mirror, highlighting the issue of data bias.

2. The Speaker's Perspective:

• The speaker identifies as an enterprise architect, a development economist, and an "African diaspora baby."
• From these perspectives, AI is seen as a potentially transformative system that could either entrench inequality or promote global equity.
• Personal anecdote about a broken ankle highlights the lack of diverse data in AI training sets.

3. Inclusion Gaps in AI:

• The African Gap: Only 1% of data used in large language models is African language data (remark made by the president of global health of the Gates Foundation at the most recent AI action summit).
• The Representation Gap: Only 20% of AI professionals globally are women, and only 1.7% are black people (Royal Society in the UK).
• The Language Gap: 90% of data used to train AI systems is in English, while only 18% of the world speaks English (research by Matina Tenz in 2023).

4. Consequences of AI Bias:

• AI shapes decisions in various domains: hiring (CV filtering), healthcare (triage), finance (credit scoring), and social media (content algorithms).
• CV Filtering Bias: University of Washington (2024) found that white-sounding names had an 85% chance of moving to the next stage, while black-sounding names had only a 9% chance.
• Facial Recognition Bias: Dr. Joy (MIT Media Lab) found error rates of 0.08% for lighter-skinned men but 34.7% for darker-skinned women.
• AI bias can affect everyone, not just specific groups, impacting hiring, healthcare, and other areas.

5. AI for Good: An Example:

• In Kirhe, Rwanda, a community hospital uses an AI-powered app to detect infections in surgical sites of post-C-section mothers.
• The app works offline and was developed by multidisciplinary teams from Harvard, MIT, and other institutions.
• This example demonstrates the potential of inclusive AI design for home-based care in low-resource settings.

6. Addressing Digital Exclusion and Data Poverty:

• Digital exclusion leads to data exclusion.
• Only 40% of people in Africa have internet access, and in the Caribbean, it can be as low as 11%.
• The focus should shift from digital exclusion to "digital hunger" and "data poverty" to increase interest in improving access.

7. Rebuilding Trust and Establishing Global Standards:

• Communities exposed to exploitation and surveillance may be hesitant to share their data.
• Global standards and policies are needed to protect, value, and respect data.
• Data breaches should be treated with the same seriousness as money laundering.

8. Data as an Economic Asset:

• Companies make billions from data, but the people who generate the data often don't share in the economic benefits.
• Companies should disclose the inclusive characteristics of their data, similar to nutritional information on food packaging.

9. Addressing the Skills Gap:

• Lack of diversity in AI is a skills gap that needs to be addressed urgently, similar to skills gaps in healthcare or hospitality.
• Governments must do more to improve representation of women and black people in the AI ecosystem.

10. Conclusion:

• Building a mirror that includes everyone means better opportunities and healthcare for all, including those in underserved communities.
• Inclusive AI can help humanity better prepare for the future.