Project summary

A practical playbook that helps AI safety advocates and policy professionals communicate effectively with the U.S. government during the short window of opportunity that may open during an AI-related crisis.

What are this project's goals? How will you achieve them?

The problem:
At the time of a major AI-related incident, like an AI-enabled cyber attack, an AI-assisted biological incident or a loss-of-control event, it's likely that the U.S. government will take fast action under pressure.
Due to the rapid progress of AI, it's likely that AI safety policies will need to be implemented in crisis conditions, but the policy literature is severely lacking in understanding how best to advocate for certain policies during these fast-moving crisis windows.
Understanding how crisis policymaking works could be critical for implementing AI safety solutions that address X risk. We need to know who to build relationships with, through which channels, with what type of messaging, and on what timelines. I have found no resources that discuss this topic as it refers to U.S. domestic policy making on AI. This project will fill the gap.

What I will do:

• Analyze five historical windows when the U.S. government implemented new policy to deal with a fast-moving crisis. For each case study, I will extract takeaways about what actions and structures had the strongest effects on which policies got put in place.

• Interview experienced AI policy advocates and congressional staff to surface the tacit knowledge of how this actually works, and where organizations feel underprepared.

Concrete output:
A written policy communication strategy playbook that contains:

1. Short breakdowns of each historical case study I looked at along with actionable takeaways

2. An guide suitable for low-capacity organizations who want to improve their policy positioning

3. A more aspirational capacity-building section that's aimed at funders and org builders to know where it might be good to put resources for policy impact at the time of crisis

4. Action plans that can be triggered at the time of crisis to make use of pre-positioning and networks

Each section will have a quick reference card containing the key insights.

I plan to publish through an established institution to increase credibility and reach.

As I work, I will be releasing a series of blog posts to stress test the arguments and quickly show progress.

Theory of Impact

AI is moving quickly, and the time between emerging publicly visible risks and the end of our window to prevent catastrophic harms might be short. Often, policy making has years or even decades to implement solutions. AI safety might have to move quicker, and there is very little literature on how to influence policy outcomes on short timelines in the midst of a crisis. AI safety policy advocates need to know how this process works to have the best possible chance of making AI governance go well.

This project applies examples of past crisis policymaking directly to AI governance to improve the strategies used by AI safety policy organizations.

Theory map:

Assumptions:

• An AI related crisis is possible

• AI safety policy might be implemented in crisis conditions.

• If I build this guide, policy advocates will change their strategies

Generally understood to be true:

• Crisis policy making functions differently than standard policy making.

• There is insufficient research on crisis policymaking

Chain of effect:

1. Improve understanding of crisis policy making in conditions similar to an AI crisis

2. AI safety advocates change their strategies based on my findings

3. Good X-risk relevant policies are more likely to be implemented at the time of a crisis

4. X-risk is reduced

Validation I've already received.

• A staff member at BlueDot Impact said the project "sounds like it could be highly impactful" and that "most people in the field acknowledge [it] matters (to at least some degree) but nobody has really done the work to systematize this."

• Staff at a leading AI safety advocacy organization said: "If it was genuinely detailed and well researched I think we would genuinely reference it in a crisis," adding that they had had the same thought themselves but lacked the time, and that they don't think the other larger AI safety advocacy groups "have a fleshed out plan like this either."

• An AI policy expert said that case studies would be useful.

• Staff at the AI governance organization whose report prompted this project said: "I think this is a valuable project ... what's most needed is past examples of influence and how it happened, but mostly the playbook that AI experts should have to be ready by the time of the crisis + the immediate aftermath."

• An independent AI policy practitioner called it "an exciting project" and offered to suggest further contacts once the draft matures.

The people I have spoken with are consistently most excited about the case studies, which suggests the historical analysis is the component with the clearest, most immediate demand.

How will this funding be used?

Minimum that makes the project happen: $20,150

• Labor: Salary: $19,500 (3 months of DC median income)

• Tools: Claude Max: $300 (3 months × $100)

• Flight to DC: $350

Ideal (fully funds the best 6 month version): $41,500

• Labor: Salary: $39,000 (6 months of DC median income)

• Tools: Claude Max: $600 (6 months × $100)

• Flight to DC: $350

• SF conference

  • Round-trip flight: $500

  • Housing (with friend): $0

  • Registration: $600

  • Meals + ground transport: $450

  • SF subtotal: $1,550

I do not currently have the financial runway to relocate to DC, but my work would be much more effective while living in D.C. I have a cheap group-house opportunity in DC that makes this feasible if funded.

Who is on your team? What's your track record on similar projects?

I am the primary person doing this work, but I'm actively recruiting collaborators. I have a bachelor's degree in science communication. I've completed the AGI Strategy course and the Frontier AI Governance course from Bluedot Impact, as well as the Aspen Policy Academy's Getting Through to Government course. I've done over 700 hours of active study on AI safety and AI policy. I also co-lead the Community Engagement Group in Utah's Responsible AI Community Consortium.
I'm a strong fit for this project because of my experience in both communication strategy as well as AI safety policy. This project is also building a base of quality outputs needed to secure other AI safety policy roles in the future.

What are the most likely causes and outcomes if this project fails?

The interview failure modes

Failure cause: I'm unable to get interviews with key policy actors.

• Effect: I don't learn who is actually involved in AI policymaking or where it happens, and the playbook directs advocates toward the wrong people and channels.

• Solution: I ask better positioned people for referrals and endorsements to enable interviews. Five people are already on record supporting this project, including staff at a leading AI safety advocacy org, staff at the AI governance org whose report prompted this work, and an independent practitioner who offered to suggest contacts. I also secure institutional affiliation early rather than at publication.

Failure cause: I get interviews, but they lack the necessary candor and clarity.

• Effect: I know who and where, but not why and how. I miss the less visible details of the policy process.

• Solution: I offer both on the record and off the record interviews, defaulting to off the record. I ask people about how other organizations handled situations rather than about their own decisions, which produces more candor. I send draft findings back to interviewees for correction, which catches incomplete accounts and builds buy in.

The adoption failure mode

Failure cause: The playbook gets distributed, but there isn't enough support for implementing its recommendations.

• Effect: The knowledge exists but gets set aside, both before a crisis and during one.

• Solution: I get agreement from someone in each organization who takes ownership of the recommendations. I run tabletop exercises, a one hour crisis simulation with each organization. Tabletop would be more feasible in the 6 month version of this project.

Overall

• A confidently wrong playbook used during a real crisis would be worse than no playbook, which is why these solutions focus on verifying findings.

How much money have you raised in the last 12 months, and from where?

I have received no funding in the last 12 months.

Grantmaking AI link:
https://app.grantmaking.ai/projects/78520cc7-cab9-4aa3-be16-60f4e741c2a2