The policy–evidence relationship
This July 2025 Policy Forum argues that evidence and governance should reinforce each other: scientific understanding should guide AI policy, while policy should make useful evidence easier to produce. It is an institutional proposal, not an empirical test of particular regulations. Political incentives, public values, and organizational constraints still shape decisions.
How should evidence inform policy?
Credible evidence varies by domain: experiments and observational data may be central in health, while economic policy also uses theory and forecasts. The authors argue for evidence that is actionable and attentive to the additional risk introduced by AI relative to existing technologies.
Evidence-based rhetoric can also justify delay or be manipulated by interested parties. Historical tobacco and fossil-fuel examples motivate careful incentive design. The paper discusses the US Evidence Act as an institutional mechanism available in its 2025 policy setting.
What is the state of the evidence?
AI evidence certification and synthesis are immature. The paper highlights the 2025 International Scientific Report on the Safety of Advanced AI [9], which combines expert judgments about originality, impact, transparency, and conflicting findings rather than relying only on peer review.
The report distinguishes malicious-use, malfunction, and systemic risks. Some harms already have direct evidence, while the magnitude of large-scale labor impacts, enhanced cyber or biological misuse, and loss of control remains uncertain. Disagreement should motivate better measurement rather than imply either certainty or the absence of risk.
How can policy accelerate evidence generation?
Policy-makers face a dilemma: premature action can be ineffective, but waiting can expose society to harm. The authors propose six mechanisms for actively improving the evidence base. These recommendations address the conditions for learning, rather than supplying a complete rule for acting on every kind of evidence.
Incentivize pre-release evaluation
Incentivize developers and trusted external evaluators to assess models before release, measuring incremental risk. Public reporting should explain evaluator independence, funding, confidentiality restrictions, testing time, access, and how findings influenced the release decision. These conditions determine how much confidence a score deserves.
Increase information sharing
Require proportionate disclosure of safety practices and evaluation results to governments and, where appropriate, the public. Journalists, researchers, and civil-society groups also identify harms, so disclosure only to regulators can leave important accountability gaps.
Obligations should match the information needed and who can use it. Thresholds should have explicit updating processes and avoid relying entirely on training compute or monetary cost as proxies for risk.
Monitor postdeployment impacts
Expand evidence about actual adoption and harms after deployment. The authors point to usage analysis and incident databases as complementary to pre-release tests. Reporting systems need clear definitions of reportable events, responsible entities, recipients, and coordinated disclosure procedures.
Cybersecurity incident reporting provides a design precedent, especially when a vulnerability affects multiple providers. The proposal is to improve ongoing observation, rather than assume laboratory evaluations capture all societal effects.
Protect third-party research
Protect good-faith independent testing of released systems. Existing access terms can inhibit evaluation when testing safeguards requires actions otherwise prohibited by providers. The paper supports a safe harbor tied to established research conduct and responsible disclosure [14], [15]. This is a policy recommendation, not a statement that such protection already applies generally.
Prioritize well-evidenced interventions
Use evidence to choose intervention points across the whole chain of potential harm. Layering model-level protections with societal defenses can reduce both AI-related and pre-existing risks. The authors use stronger software security and controls elsewhere in biological-material supply chains as examples; restricting a model is only one possible intervention.
Bridge fragmented subcommunities
Support deliberative processes that clarify both agreement and uncertainty across fragmented research communities. Scientific panels and scientist–policy-maker partnerships can synthesize evidence, provided participants have relevant expertise, broad inclusion, and independence from undue pressure. Consensus should remain credible within the scientific community and to outside stakeholders.
A long-term governance challenge
Evidence-generation mechanisms need sustained scientific research, including work toward safer technology. Future effort should also connect evidence to action through conditional response plans: parties can agree on what to do if a finding appears even when they disagree on its likelihood or timing.
An evidence-based approach does not remove democratic disagreement about goals and trade-offs. It supplies a shared factual basis for that debate. The paper’s institutional examples and assessments describe its July 2025 publication context.
Acknowledgments and disclosures
The authors thank colleagues for discussion and state that the article reflects their personal scholarly views, not their institutions’ positions.
Disclosed affiliations include Arora’s part-time Meta employment; Choi’s NVIDIA consultancy and former Ai2 consultancy; Cuéllar’s Stanford advisory and visiting roles and Inflection AI board membership; Fei-Fei’s World Labs role; Koyejo’s Virtue AI cofounding role; Nelson’s Mozilla, Oxford AI-ethics, and Data and Society board/advisory roles; Venkatasubramanian’s Data and Society and Partnership on AI boards; Stoica’s executive-chair roles at Databricks and Anyscale; Liang’s Together AI, Simile AI, Arell AI, and Virtue AI roles and Snowcrash/Humanitas advising; and Song’s Oasis Labs founding and Virtue AI role.