Why Science Matters in Global AI Governance

Contents

Executive Summary

This AI summit session emphasizes that effective global AI governance requires a science-based foundation rather than guesswork or hype. The establishment of the UN's Independent International Scientific Panel on AI—designed to provide shared, evidence-based analysis—is positioned as critical infrastructure for bridging the gap between rapidly advancing AI capabilities and policy-making that protects people while enabling innovation.

Key Takeaways

Science is not optional for AI governance — It provides the shared factual baseline needed to move beyond philosophical debate to technical coordination, reducing fragmentation and enabling interoperability across jurisdictions.
Speed and care are not mutually exclusive — High-level principles paired with technology-enabled guardrails can allow rapid innovation while embedding safety through design rather than reactive regulation.
The Global South cannot be an afterthought — Equitable governance requires centering voices, evidence, and impacts from developing countries and marginalized populations; one-size-fits-all policies will fail.
Uncertainty is not an excuse for inaction — Risks with catastrophic potential demand policy attention even without complete proof; waiting for certainty is itself a policy choice with consequences.
Institutions matter — The UN's unique legitimacy and inclusiveness make it essential infrastructure for facilitating global AI governance; fragmentation across multiple initiatives risks safety and equity.

Key Topics Covered

Science as foundation for AI governance — The principle that policy cannot be built on hype or disinformation; facts and evidence must guide decisions
The UN Independent International Scientific Panel on AI — Its mandate, composition, multidisciplinary approach, and role in closing the "AI knowledge gap"
The science-policy interface — How scientific evidence translates into actionable policy; challenges of rapid technological change outpacing research timelines
Global South perspectives — Ensuring equitable representation and addressing context-specific impacts of AI in developing countries
Risk assessment and uncertainty — Addressing AI risks even when certainty is unavailable; tipping points and catastrophic scenarios
Responsible innovation — Balancing speed with care; embedding governance through technical design
Interoperability and standardization — Need for shared benchmarks, testing methodologies, and technical standards across jurisdictions
Human agency and oversight — Making human control a technical reality; accountability and transparency in high-stakes decisions
AI for sustainable development — Harnessing AI for health, education, agriculture, and other SDGs; building evidence of real-world impact
Capacity building — Ensuring all countries, especially small states and Global South nations, can meaningfully engage with AI governance

Key Points & Insights

The Knowledge Gap Problem: Rapid AI innovation is "moving at the speed of light," outpacing collective ability to understand and govern it. Policy lags behind capability development, creating systemic risk.
Science-Based Governance as Accelerator, Not Brake: Contrary to industry concerns, science-led governance strengthens rather than inhibits innovation by enabling smart, risk-based guardrails instead of blunt regulatory instruments.
Universal Language Through Science: Science can serve as a common baseline across fragmented regulatory regimes—aligning technical testing, measurement, and safety standards across regions prevents a costly "patchwork" of incompatible rules.
Uncertainty and Risk Severity: Even without proof of catastrophic outcomes, policy-makers must attend to risks with high severity potential (e.g., AI tipping points) and low certainty. The climate-change analogy applies: we cannot wait for complete evidence before acting.
Uneven AI Capabilities Create Governance Complexity: AI systems surpass humans on some capability measures while performing at age-6 level on others—this uneven development makes assessment and prediction difficult and requires granular, context-aware evaluation.
The Lag Problem in Research and Policy: Scientific papers take time to publish; studies involving humans take months; policy discussions happen even later. By the time strong evidence emerges, technology may have already moved further, making proactive governance essential.
Context Matters for Impact: AI effects are not globally uniform. Impacts on low-income farmers in India differ from effects on mechanized agriculture in Europe; interventions in high-income countries may not translate to low-income contexts. Evidence must be locally grounded.
Computational Efficiency Has Systemic Implications: Example: 90% of AI is matrix multiplication—so even 1% efficiency gains have massive energy implications. Detailed technical understanding is essential for policy-makers.
Evidence Gaps in Critical Domains: Even in promising areas like AI in education, causality remains unclear (does more AI use cause better outcomes, or vice versa?). Robust empirical research is lacking across Global South contexts.
Trust in Science Depends on Leadership: During COVID, countries led by scientists using iterative, evidence-based policy achieved better outcomes. Similarly, AI governance requires institutions and leaders willing to adapt policy as evidence emerges.

Notable Quotes or Statements

"We cannot govern what we do not understand."
— Opening premise of the session

"AI does not stop at borders and no nation can fully grasp its implications on its own."
— UN Secretary General

"Less noise, more knowledge."
— UN Secretary General

"Science-led governance is not a break on progress. It is an accelerator for solutions, a way to make progress safer, fairer, and more widely shared."
— UN Secretary General

"Science informs but humans decide. Our goal is to make human control a technical reality, not a slogan."
— UN Secretary General

"The average grade [for tech industry predictions over 30 years] was 25%. You couldn't even get close to the top of failing."
— Brad Smith, Microsoft

"Human capability is neither fixed nor finite."
— Brad Smith, Microsoft

"Nothing in life is to be feared. Everything is to be understood."
— Marie Curie (quoted by Anne Bubbero)

"If AI has to work for everyone then we need to make sure that those voices are heard."
— Somia (WHO)

"Even in something as simple as education... we need a lot more evidence to come in place."
— Prof. Balaraman Raindran

"Scientific input [should be viewed] not as a constraint on policy flexibility but as a foundation for more durable effective governance that can maintain public trust."
— Minister Josephine Teo, Singapore

Speakers & Organizations Mentioned

Speaker	Role/Organization
UN Secretary General Antonio Guterres	Opening remarks on UN role in AI governance
Prof. Yoshua Bengio	Scientific Director, Mila; Most-cited living computer scientist; Member, UN Independent Scientific Panel on AI
Brad Smith	Vice Chair & President, Microsoft Corporation
Somia	Former Chief Scientist, WHO; Expert on science-policy interface during COVID-19
Prof. Balaraman Raindran	AI researcher, IIT Madras; Member, UN Independent Scientific Panel on AI
Ajay Kumar Sud	Principal Scientific Adviser to the Government of India
Anne Bubbero	France's Special Envoy for AI
Amandeep Singh Gil	UN Under-Secretary-General & Special Envoy for Digital & Emerging Technologies (Moderator)
Minister Josephine Teo	Minister for Digital Development & Information, Singapore

Key Institutions Referenced:

United Nations (UN) and General Assembly
WHO (World Health Organization)
IPCC (Intergovernmental Panel on Climate Change) — cited as model
ILO (International Labor Organization)
Microsoft Corporation
IIT Madras
Singapore's designated AI Safety Institute
ASEAN (Association of Southeast Asian Nations)

Technical Concepts & Resources

Concept/Tool	Context
Independent International Scientific Panel on AI	UN-mandated body with 40 globally diverse, multidisciplinary experts; designed to provide shared baseline of analysis and close the "AI knowledge gap"
Benchmarks and standardized evaluation methodologies	Critical need for shared, interoperable testing standards to allow consistent assessment of AI capabilities and risks across jurisdictions
Matrix multiplication efficiency	Example: AI systems rely heavily (90%) on matrix multiplication, so even 1% efficiency gains have major systemic energy implications
Tipping points analogy	Risk management framework borrowed from climate science: attending to low-probability, high-severity scenarios even without complete evidence
Technical design for governance (Technolegal governance)	India's approach: embedding governance constraints through technical architecture rather than purely regulatory rules; informed by digital public infrastructure (Aadhaar, UPI) deployment experience
Red teaming and safety testing	Singapore's AI Safety Red Teaming Challenge; multilingual and multicultural exercises; international network for advanced AI measurement, evaluation, and science
Scientific benchmarks tracking rapid capability growth	Evidence of uneven capability development: systems surpass humans on some measures while performing at age-6 level on others
Multidisciplinary assessment frameworks	Need for integration of machine learning, applied AI, social science, ethics, economics, health, education, labor, agriculture perspectives
Global AI safety research priorities	Singapore Consensus; ASEAN Guide on AI Governance and Ethics; evolving tools for AI safety testing and benchmarking

Context & Event Details

Event: AI Summit in Delhi (implicitly a global governance dialogue on AI)
Date: Second month of second quarter of 21st century (~April 2024, based on references to recent events)
Scope: Global governance with emphasis on UN role, international scientific cooperation, and inclusion of Global South perspectives
Related UN Initiatives: High-Level Advisory Body on AI (report: "Governing AI for Humanity," end of 2024); Global Dialogue on AI Governance (scheduled for July)

[End of Summary]