Governing AI: Risks, Freedoms, and Democratic Oversight

Symposium Overview

• 💡 The symposium focuses on artificial intelligence and democratic freedoms, exploring challenges and interventions to mitigate harms and leverage AI for societal benefit.
• 🧠 It emphasizes an interdisciplinary approach, bringing together experts from computer science, philosophy, political science, economics, and law.

AI Risk Governance Challenges

• ⚠️ Atoosa Kasirzadeh discussed the challenges of catastrophic AI risk governance, highlighting limitations of current safety frameworks and the need to understand risk dimensions like source, causal mechanisms, and impact.
• 🎯 Deirdre Mulligan advocated for a sociotechnical, systems-level approach to AI risk management, moving beyond model-centric governance and distinguishing between hazards and actual harms.
• 🔑 Mulligan stressed the importance of enlisting diverse stakeholders and expertise for effective risk mitigation, as asset owners may lack the necessary capacity or legitimacy.

AI and Democratic Institutions

• 🏛️ Alondra Nelson presented the AI Bill of Rights as a "civic architecture," fostering public engagement in governance and serving as a framework for state legislation and educational curricula.
• 📰 Arvind Narayanan warned about the weakening of journalism by AI, exacerbated by power asymmetry in deals between AI companies and news publishers, suggesting taxing AI companies to fund journalism.
• 🔒 Concerns were raised about the invisible surveillance ecosystem in democracies, with government access to vast data eroding privacy and freedom of speech.

Rethinking AI Evaluation

• 🔬 Panelists highlighted the inadequacy of current AI evaluations, noting a lack of access to crucial information (training data, parameters) from powerful systems.
• ✅ Deirdre Mulligan argued for focusing on real-world impact through in-situational testing involving social scientists and practitioners, rather than solely lab-based model evaluations.
• 📈 Arvind Narayanan emphasized the need for rigorous, documented real-world use cases to understand AI's actual effects, similar to how cyber security learns from "in the wild" observations.

Future of AI and Regulation

• ⚖️ Discussions touched on the EU AI Act as a model for policy development and the need for sectoral regulation with broad requirements like transparency.
• 🤝 The panel underscored the importance of resisting polarization in AI discourse and fostering civil conversations to find common ground on AI's future and its governance.