Top Scholars Call for Evidence-Based Approach to AI Policy
Date: July 31, 2025
Topics: Regulation, Policy, Governance
Experts from Stanford HAI and leading universities urge policymakers to ground AI governance in scientific evidence. As artificial intelligence continues to influence industries and society, developing thoughtful and effective policy has become crucial. The question is: how can policymakers best leverage scientific research to govern this fast-moving field?
Scholars from Stanford, UC Berkeley, Princeton, the Carnegie Endowment for International Peace, and others advocate for an evidence-based approach to AI policy. A recent paper published in Science outlines their recommendations. This paper includes contributions from prominent researchers such as Fei-Fei Li, Yejin Choi, Daniel E. Ho, Percy Liang, and Rishi Bommasani of Stanford’s Institute for Human-Centered AI.
What does the paper address?
The paper offers a clear vision for integrating evidence into AI policymaking. While the idea of evidence-based policy is widely accepted in principle, applying it effectively in the AI domain is complex. The authors highlight how policymakers can accelerate evidence generation rather than wait for perfect information, which could otherwise stall governance efforts.
What is meant by evidence-based policy?
In theory, good policy should incorporate evidence grounded in scientific understanding. However, what counts as evidence varies across fields. For example, health policy often relies on randomized controlled trials, while economic policy may depend on macroeconomic forecasts. The paper stresses the need to clarify what types of evidence are suitable for AI governance.
Challenges to adopting this approach
Identifying credible evidence is a primary challenge. Different domains accept different standards—some prioritize observed outcomes, others theoretical models. Additionally, there is the risk that demands for evidence can be used to delay or obstruct regulation. Partial or selective evidence might be presented to question scientific certainty and impede policy progress.
What can policymakers do today?
The paper emphasizes how policymakers can encourage the creation of credible evidence. Key recommendations include:
- Increasing transparency from frontier AI companies on critical issues such as risk mitigation.
- Incentivizing external, independent testing of AI models before public release, with published results and clear documentation of testing conditions.
- Establishing safe harbors to protect third-party evaluators conducting safety assessments in good faith.
Beyond evidence generation, policymakers must develop mechanisms to synthesize and interpret the growing volume of AI research and reports. Initiatives like the International Scientific Report on the Safety of Advanced AI demonstrate how expert consensus can inform policy. However, fostering scientific consensus remains difficult due to divergent views within the AI community.
Policymakers can help create environments that encourage serious attempts at consensus-building. Even acknowledging uncertainties transparently would strengthen the foundation for better AI governance.
For those working in science and research who are interested in deepening their expertise on AI policy and governance, exploring courses on Complete AI Training can provide practical knowledge aligned with current developments.
Further reading: “Advancing Science- and Evidence-Based AI Policy,” Science, by Rishi Bommasani et al.
Your membership also unlocks:
