Foundations of Evidence-Based AI Policy with Rishi Bommasani

The Need for Evidence-Based AI Policy

• 💡 AI's profound impact necessitates robust governance frameworks to produce better societal outcomes.
• ⚠️ Current policy-making often lacks scientific rigor, prioritizing public engagement over evidence validity.
• 🎯 Rishi Bommasani's work aims to establish a scientific subfield for AI policy, bridging computer science with other disciplines.

Holistic AI Measurement Frameworks

• 🔬 Measurement is foundational for understanding AI's societal impact and informing policy design.
• 🔑 Helm (Holistic Evaluation of Language Models) provides third-party, standardized, and continuous assessment of AI models, addressing gaps in developer-led evaluations.
• 📊 Research extends beyond models to deployed AI systems (e.g., hiring algorithms) and AI companies, analyzing their real-world effects and transparency.

Uncovering Bias in AI Systems

• 🔍 Analysis of hiring AI systems (like Pimetrics) reveals pervasive use and significant societal impact.
• ⚖️ Despite vendor claims, disaggregated data shows significant bias against certain racial groups in specific job positions, meeting legal standards for discrimination.
• 🧩 The widespread use of single AI vendors leads to algorithmic monoculture and increased homogeneity in hiring outcomes, potentially prolonging unemployment.

Measuring AI Company Transparency

• 📈 A transparency index evaluates major AI companies across their supply chain, from data acquisition to deployment impact.
• 💡 Findings reveal significant opacity across the industry, particularly concerning data, compute, and downstream use.
• 🚀 The index acts as an incentive mechanism, encouraging companies (especially smaller ones) to increase disclosures and fostering a more transparent ecosystem.

Informing Global AI Policy

• ✅ Research provides conceptual frameworks (e.g., "foundation models") adopted by governments like the US, EU, and California.
• 🤝 Direct engagement with policymakers, such as assessing the marginal risk of open AI models for the US government and informing the EU AI Act.
• 📜 The California Report on Frontier AI Policy directly influenced new state laws, demonstrating how strong technical foundations can build policy consensus.