兰德-人工智能政策项目（AIPP）召开-汇集技术和政策专家_利用人工智能改善政策制定（英）-2025_19页_220kb

AI-Enabled Policy Project (AIPP) Convening Summary

Core Content

The AI-Enabled Policy Project (AIPP) hosted a one-day working session in Washington, D.C., on May 15, 2025, to explore how artificial intelligence (AI) can support various components of the policymaking life cycle, from research to formulation to stakeholder engagement. The event was co-hosted by RAND, the Stimson Center, and the Tony Blair Institute for Global Change, and attracted 56 participants from diverse backgrounds, including policy researchers, policymakers, and technology experts.

The workshop was structured into four sessions, each with guided questions and small group discussions. Each session concluded with a share-out, where representatives from each table presented key insights to the larger group. The session notes were used to synthesize the overall outcomes and insights of the event.

Main Sessions and Key Insights

Session 1: Defining AI-Enabled Policymaking

• Objective: Define the scope of AI-enabled policymaking by mapping out the policymaking workflow.
• Key Components Identified:
  • Early-stage policy activities such as monitoring media, gathering official statements, and identifying emerging issues.
  • Stakeholder engagement tasks like public opinion polling, sentiment tracking, and feedback solicitation.
  • Analysis tasks such as cost-benefit analysis, counterfactual analysis, and historical analysis.
  • Additional components include coalition building, awareness and education for new policies, and agenda setting.
• Observation: The policymaking process is often nonlinear and iterative, which may be difficult to capture in a linear framework.

Session 2: Core AI Capabilities and System Requirements

• Objective: Explore AI's existing and potential applications in policymaking, including tool capabilities, automation opportunities, and data flow considerations.
• Current AI Tools Used:
  • General-purpose large language models (e.g., ChatGPT, Claude) for background research, literature reviews, and project planning.
  • Tools like Microsoft Copilot, Harvey, and Policy Synth for specific tasks such as legislative drafting, legal workflows, and policy synthesis.
• Challenges Identified:
  • Privacy and legal concerns.
  • Inconsistent trust in AI outputs.
  • Limited access to nonpublic or tacit knowledge.
  • Data scarcity and gaps in existing datasets.
• Opportunities:
  • Automating routine and formulaic tasks.
  • Enhancing stakeholder engagement and feedback loops.
  • Customizing AI systems for specific user needs.

Session 3: Implementation Considerations

• Objective: Address trustworthiness, democratic alignment, and risks of AI integration into policymaking.
• Key Concerns:
  • Erosion of human agency due to over-delegation to AI.
  • Information overload and uncritical acceptance of AI outputs.
  • Bias and ideological assumptions embedded in models.
  • Cognitive decline and loss of judgment among policymakers.
• Recommendations:
  • Meaningful feedback loops to ensure models are contextually grounded.
  • Human calibration and oversight to interpret data correctly.
  • Verification mechanisms to reduce noise and spam in public feedback.
  • Proactive policy interventions to address accountability and consumer protection gaps.

Session 4: Building the Technical Roadmap

• Objective: Map existing AI tools to policy applications and define a technical roadmap for future AI-enabled policymaking.
• Existing Tools:
  • General-purpose models (e.g., OpenAI, Perplexity) for research and drafting.
  • Specialized tools (e.g., Elicit, Policy Synth) for literature reviews and policy synthesis.
  • Tools like Granola, Grain, and Microsoft Copilot for meeting capture and summaries.
• Desired New Tools:
  • Custom workflows for generating policy summaries, talking points, and audience-specific outputs.
  • Tools that convert meeting notes into structured formats (e.g., relationship maps, CRM entries).
  • AI systems that translate complex rules into plain language.
  • Automated analysis tools for policy alignment and fact-checking.
  • Platforms to enhance public engagement and transparency.
• Deployment Challenges:
  • Lack of financial incentives for companies to develop government-specific tools.
  • Need for reliable user interfaces and training programs.
  • Multiyear contracts, incentive design, and viable business models to justify development.

Key Information

• Participants: 56 individuals from diverse backgrounds in policy and technology.
• Organizers: RAND, the Stimson Center, and the Tony Blair Institute for Global Change.
• Funding Sources: Income from operations, gifts from philanthropic supporters, and contributions from various foundations and organizations.
• Attendee Organizations: Includes entities like RAND, the Stimson Center, MIT GOV/LAB, and the Future of Life Institute.
• Technical Roadmap Focus: Mapping AI tools to specific policy tasks, improving system reliability, and ensuring democratic alignment and transparency.
• Challenges:
  • Privacy and legal concerns.
  • Data access and quality.
  • Trust in AI outputs.
  • Integration into existing workflows.
  • Potential for bias and ideological reinforcement.
• Opportunities:
  • Automation of routine tasks.
  • Enhanced stakeholder engagement.
  • Democratization of access to high-quality analysis.
  • Faster feedback loops for policy implementation.

Conclusion

The AIPP convening aimed to bring together technology and policy experts to explore the role of AI in improving policymaking. The event provided a platform for identifying current capabilities, future opportunities, and implementation challenges. A key takeaway was the need for human oversight, transparency, and ethical considerations in AI-enabled policymaking. The session also emphasized the importance of building AI systems that are aligned with democratic values and addressing the limitations of general-purpose models. The final outcome was a technical roadmap that outlines the next steps for developing and deploying AI tools in the policymaking process.