The public's lack of trust in government and the continuing rise of AI are two persistent stories in the news this year. In a March 2025 survey (PDF), trust in the U.S. government was down to 4 percent and only 11 percent said America is a mostly fair society. In another survey less than 40 percent agreed that the government has a positive impact on people like them. This amounts to a political efficacy crisis, with too many feeling the government is not responsive to their needs and they cannot meaningfully participate in making governance better.
Might the second story, the rise of AI, help address the first? Could AI be leveraged to empower the public by making the creation of policy more open and participatory? Are there some simple ideas for prompts that could help? We conducted a small independent research project to explore these questions.
A first step is to consider the nature of policy “problems.” Rather than trying to get a single right answer, we sought policy recommendations that would satisfy a complex set of general policy objectives. We took into account resource constraints, a range of specific stakeholder needs, and risk reduction requirements. We represented these using direct in-prompt descriptions combined with descriptions of virtual stakeholders who can be queried to review a policy or make recommendations.
Metaheuristics are algorithms that search through sets of incremental options for improvement, seeking to move to new policy recommendations that are systematically better than the current version.
The second step was to think about how policy recommendations could be improved to better align with objectives, resource limitations, specific needs, and risk reduction. We developed prompt texts that emulated the concepts of metaheuristics from the field of operations research (OR). Metaheuristics are algorithms that search through sets of incremental options for improvement, seeking to move to new policy recommendations that are systematically better than the current version.
We crafted a metaheuristic in a prompt by having virtual stakeholders evaluate the current proposals with respect to their key concerns and suggest improvements. We then had the LLM accept ideas that improve the overall satisfaction of the requirements and prepare a new set of policy provisions. These new provisions were then evaluated by the virtual stakeholders again and so on.
For example, we experimented with this approach to generate a violent crime reduction strategy, and the resulting draft was substantially more detailed and aligned with evidence than simply asking the LLM for a strategy. We captured results in a new RAND perspective, “The Well-Tempered AI Assistant for Policy Processes,” a part of RAND's Social and Economic Policy Rethink Initiative.
To be sure, these drafts are suited only to start discussion within a larger human-driven policymaking process. However, the outputs were better for starting discussion than, say, a blank page or a status quo. Much could be done to improve on this very small-scale pilot effort. Results could be improved by using more sophisticated heuristics and other OR techniques, providing more detailed treatment of the virtual stakeholders and other prompting techniques, providing more specialized knowledge to work from, or employing an agentic AI approach.
Moving forward, OR-enhanced AI could expedite and align the full policy creation process.
Moving forward, OR-enhanced AI could expedite and align the full policy creation process. In addition to helping generate and evaluate options, AI shows promise in summarizing and synthesizing large-scale feedback from many stakeholders—including via open public comment—and subject matter experts. This could make possible large-scale policy generation and revision events.
Many times, policy forums and workshops offer participants limited to no role in crafting outputs. AI could change that, helping provide better discussion drafts and facilitating real-time policy inputs and revisions during discussion sessions. In the longer term, AI could help empower the public by facilitating widespread and deep participation in policy and planning.
Combining human expertise, AI, OR methods, and large-scale feedback holds promise for enhancing governance and public trust. Further research and collaboration is needed to advance the use of AI and OR to empower people to make better policies and plans.
