How do we make AI truly inclusive? | On Second Thought

Key Concepts

• Fairness in AI: Allocation harms, quality of service harms, and representational harms.
• Responsible AI Standard: Microsoft’s foundational playbook for developing AI responsibly, structured by AI principles, goals, and requirements.
• Global Perspectives AI Fellowship: A program designed to incorporate diverse perspectives into AI policy development.
• Digital Inclusion: Addressing barriers to online access and ensuring AI solutions are relevant to local contexts.
• Representational Harms: AI systems over/under representing or erasing certain groups, or perpetuating stereotypes.
• Transparency in AI: Disclosing AI system usage, limitations, and intended use to developers and users.
• Cultural Context Awareness: Designing AI systems that understand and respect local customs, lived experiences, and nuances.
• Low Resource Languages: Languages with limited digital representation and data, requiring specific attention in AI development.

AI: Building a More Inclusive Future - A Discussion with Hiwot Tesfaye of Microsoft Responsible AI

This conversation with Hiwot Tesfaye, a Technical Advisor in Microsoft’s Office of Responsible AI, centers on the critical need for inclusivity, transparency, and global representation in the development and deployment of artificial intelligence. The discussion highlights the challenges of building AI systems that benefit all of humanity, not just those with the loudest voices or the most resources.

The Need for Diverse Perspectives in AI Development

Tesfaye emphasizes that building AI is akin to rebuilding society, necessitating broad participation in decision-making. She argues that the process must extend beyond AI engineers and technologists to include linguists, sociologists, anthropologists, and importantly, the average person. Her own background – growing up in Ethiopia, Sudan, Kenya, and Uganda – has instilled in her a sensitivity to missing perspectives and cultural nuances, shaping her approach to responsible AI. She notes her role as a “translator” between technical teams and policy makers, leveraging her data science training and cross-cultural experience. As she states, “There’s no way that [her] background doesn’t impact the way [she] see[s] the world and how [she] approach[es] [her] work at Microsoft.”

Defining and Operationalizing Fairness in AI

The conversation delves into the complexities of “fairness” in AI, outlining three key categories of fairness-related harms identified at Microsoft:

1. Allocation Harms: Disparate outcomes in consequential decisions (housing, finance) across different groups.
2. Quality of Service Harms: Variations in performance (e.g., speech-to-text accuracy) across different demographic groups.
3. Representational Harms: Biased or incomplete representation of groups in AI-generated content (images, search results), including stereotypes or erasure.

Categorizing these harms is crucial for identifying who is being harmed, how, and developing mitigation strategies. Tesfaye highlights that simply providing access to technology isn’t enough; the solutions must be contextually relevant. She cites a Microsoft Research Nairobi example where a chatbot advised a resident of Kibera (a Nairobi slum) to seek medical attention at night, failing to account for the safety risks of traveling in that area.

The Global Perspectives AI Fellowship and the Responsible AI Standard

Microsoft’s Global Perspectives AI Fellowship was created to address blind spots in AI policy development and ensure representation from underrepresented regions. The fellowship brings together AI experts, policymakers, and creatives to identify beneficial uses and potential misuses of AI globally. Currently, the fellowship is providing feedback on Microsoft’s Responsible AI Standard, a foundational document outlining principles, goals, and requirements for responsible AI development. This feedback will directly influence how AI systems and models are built and tested across Microsoft. Testing for cultural sensitivity is a significant challenge, currently relying heavily on human evaluators for languages beyond English, until generative AI systems improve in those languages.

Addressing Digital Inclusion and Language Barriers

Tesfaye stresses the importance of digital inclusion, recognizing that barriers to access extend beyond physical connectivity. She points out that even when access exists, solutions may not be relevant or useful to specific communities. In the context of AI, accessibility in local languages is paramount. She illustrates this with the example of Microsoft Teams’ live transcription feature, which, while imperfect, allows users to follow conversations in languages they don’t fully understand.

She emphasizes that language itself is a technology, constantly evolving, and AI systems must adapt to these changes. She notes that “language is not static. It changes and evolves like the next generation has a different way of speaking and engaging with one another.”

Global Governance and Regional Coalitions

The discussion acknowledges the disparity in resources dedicated to AI development globally, with investment often exceeding the GDP of entire countries. Despite this, Tesfaye argues for the inclusion of diverse voices in global AI governance. She highlights the power of regional coalitions, using the example of UNESCO’s partnership with Iceland to support low-resource languages. Iceland, facing the potential loss of its language, has proactively collected data and shared it with AI companies like OpenAI to improve model performance for Icelandic. This model could be replicated by other countries with limited language representation.

The Future of AI: Impact and Transparency

Tesfaye expresses a desire for AI to solve impactful problems, moving beyond marginal productivity gains. She is particularly excited about the potential of AI to break down communication barriers within countries with multiple languages, citing Ethiopia as an example with its 80 languages spoken across 130 million people.

Transparency is identified as a crucial aspect of responsible AI. This includes disclosing that a system is AI-powered and communicating its limitations to developers and users. Tesfaye emphasizes that AI development is an ongoing process, requiring continuous refinement and updates. She states, “we’re not done. So we’re going to continue to refine and to update how we build this technology, how we safeguard this technology.”

Conclusion

The conversation with Hiwot Tesfaye underscores the critical importance of proactive inclusivity, cultural sensitivity, and transparency in the development of AI. Microsoft’s efforts, through initiatives like the Global Perspectives AI Fellowship and the Responsible AI Standard, demonstrate a commitment to building AI systems that benefit all of humanity, acknowledging that this is an ongoing process requiring continuous learning and adaptation. The key takeaway is that building a truly responsible and impactful AI future requires a collaborative effort, incorporating diverse perspectives and prioritizing equitable outcomes.