Reproducibility: The New Frontier in AI Governance

Reproducibility problems undermine AI policy decisions: The paper argues that weak scientific standards and poor reproducibility in AI research create a "low Signal-To-Noise Ratio" that makes it harder for policymakers to understand which AI risks actually matter, potentially leading to ineffective or misguided regulations.

AI research is moving too fast for proper scientific rigor: The combination of rapid publication speeds in AI and lack of strong reproducibility protocols means findings often aren't verified before influencing policy decisions, similar to reproducibility crises that have affected other scientific fields.

Specific reproducibility practices could improve governance: The paper recommends that AI research adopt three key practices from other scientific domains: preregistration (declaring studies in advance), increased statistical power (larger, more robust studies), and publication of negative results—all of which could help policymakers build better consensus on AI risks.

Current conditions favor regulatory capture: When policymakers lack reliable information about AI risks due to reproducibility issues, they become more vulnerable to influence from well-resourced actors, potentially leading to regulations that serve industry interests rather than public safety.

Reproducibility should be a core governance tool: Rather than viewing reproducibility as just a scientific best practice, the paper argues governments should actively demand higher reproducibility standards from AI research as a fundamental part of their governance toolkit.