You should start a local org to lobby your representatives about AI x-risk

Epistemic status: early results are promising but so far this is n=1. Given the minimal time commitment required, however, I believe it's worth it for others to try this approach.

Bernie Sanders just sent a letter to OpenAI, Anthropic, and Meta calling for them to pause frontier model development. It's unlikely any of them will do this unilaterally but it’s encouraging to see a US Senator taking bold action on x-risk. To achieve what's needed to coordinate the world's frontier AI labs—an international treaty—we need many more legislators to speak out and act on this issue. 

For that reason, in February I started the Pittsburgh AI Safety Coalition (PAISC). The theory of change is to demonstrate that a broad range of constituents care about AI x-risk and to use that power to lobby our Congressional representatives to work towards an international treaty. This has taken up a relatively small amount of my free time (2-6 hours/week) and so far I’ve developed an ongoing relationship with one of our House representatives' staff (I’m in the process of reaching out to other offices as well). 

Due to the large number of topics that any given Congressional legislative assistant (LA) has to cover, it’s hard for them to stay up to date on AI risk, especially given how fast everything is moving now. That problem will only get worse. Therefore, having access to and relationships with these staffers will be crucial  for ensuring they have the necessary information to make good decisions about AI.

Given the time it takes relative to the results so far, I think this is a worthwhile strategy for others to try in their own cities. I want to share some of my experience to make those efforts more successful. Knowing what I know now, I think I could go from starting the coalition to a first meeting with a representative's office in about a month. Given the increasing pace of AI development, this rapid turnaround is another good reason to try this approach—it can have an impact in the near future.

The playbook is simple:

• Build a small coalition (of orgs or individuals)
• Schedule/hold a first meeting with your representative(s)
• Maintain a relationship

Build a coalition 

I went into this thinking I would need a membership of at least fifteen local orgs to be taken seriously. That was way off. Although I had some momentum early on building the coalition, with the League of Women Voters of Greater Pittsburgh and the Pittsburgh Musician’s union joining , that momentum quickly slowed. I’m not exactly sure why. But it was clear from the first meeting I had with a legislative assistant (LA) that even our small coalition was enough for a reasonable amount of credibility. I was encouraged by some AI safety advocates outside Pittsburgh to start lobbying even though the coalition is small, and that was the right call. 

Early on I wrote an internal strategy document that I share with prospective members. It’s a simple overview of what the coalition is, our mission and priorities, and the organizational structure. I’d be happy to share this document with anyone upon request.

Some tips for getting the first few orgs to join:

• Start with orgs with whom you have a personal connection. I am a musician so it was not hard to get a meeting with the board of the Musician’s Union. I also had a personal connection to a board member of the League of Women Voters which provided an opening.
• Cold emails basically don’t work. It doesn’t take that much time to send them out though, so it’s probably worth it on the off chance someone bites. But don’t be discouraged if no one replies. Focus primarily on organizations within your network.
• I wrote an op-ed in a local paper after the Mythos announcement in April that helped with credibility. Depending on where you live this will be easier or harder, but I think it’s worth trying early in the process. I was able to get published because the article was pegged to a recent AI issue in the news; fortunately (or unfortunately) these are happening more and more often, so an opportunity will likely present itself.
• The hardest part of building this coalition has been keeping x-risk as the top priority while still honoring AI concerns that are more salient to the target organization. I usually begin presentations with near-term threats that people are familiar with (potential for job loss, surveillance, deep fakes, etc) and then proceed  power concentration and loss of control risk. This has worked well and even organizations that ultimately decide not to join the coalition have still been receptive to and appreciative of hearing about x-risk.
   

Building a simple website and listing your membership is enough to demonstrate credibility. Ideally, coalition membership will grow and the website will evolve into a strong signal that many people in your city care about AI x-risk.

Schedule a first meeting

Call your local district office and schedule an in person meeting (these make a better first impression than virtual). If you’re meeting with a Senator, the AI staffer is going to be in DC; use the district office meeting as an intro and get the contact info for the DC staffer, with whom you can schedule a virtual meeting. Meetings can very often be scheduled within a week. That said, it sometimes takes a while to get a response from certain offices (I’m having this issue with one local office right now). Usually this is just because they’re busy, not because they’re avoiding you. It’s okay to keep being persistent until you hear back.

Different offices will have different baseline awareness of x-risk as an issue. Be prepared to explain the basics and refer them to some long-form materials that go into greater detail. The first staffer I met with was already aware of and concerned about x-risk and told me he’d been urging “the boss” to take it more seriously. The Congressperson repeatedly responded that he needs to see that x-risk is a top priority of his constituents. Great—that’s exactly what these orgs exist to do. 

The two policy actions I recommended in the meeting were 1) Work towards an international treaty. I supplied the MIRI treaty as a model but, as a person with a non-technical background, stressed that the details would need to be worked out with policy experts. 2) Introduce a Sense of the House/Senate resolution saying something to the effect of “The U.S. should coordinate internationally and work towards a treaty governing frontier AI.” While this is a non-binding resolution, it could be a good signal to China that the U.S. is willing to coordinate and give clarity to the AI safety movement which politicians are willing to coordinate and which are not.

While offering some high-level policy proposals is probably good, I think the biggest contribution from this style of non-technical local org is to build relationships with their representatives, demonstrate that constituents care about the issue, and supply a steady stream of updates from the AI world to legislative staff.

Leave behind a one-pager with a basic overview of AI x-risk and make sure to get the staffer’s contact info. 

Maintain A Relationship

Having had my first meeting before Hugging Face, I was well positioned to funnel good coverage and analysis of the incident to the legislator’s office. This prompted a meeting with the legislative assistant who covers AI where we went into more depth on x-risk. This LA has been very appreciative of receiving updates and is eager to maintain a relationship. As crazier things start happening, it will be very valuable to be able to quickly notify legislative offices with coverage that accurately represents the situation.

Mistakes

Here is a list of mistakes I made so you can avoid them. Again, I think if I hadn’t made these mistakes, the time from starting the org to a first meeting with an LA would be around one month.

• Waiting too long to start lobbying. If I did it again, I would have started as soon as I had three member orgs.
• I think it’s important that x-risk be the focus but I also know that it doesn’t resonate with a lot of already-established civic groups that have their own priorities. I’m not sure how, but I probably should have leaned more into concerns that those orgs already have without abandoning x-risk. (As an aside: I recently did the coalition organizing for Stop The AI Race's July 11th protest in San Francisco. I was able to get two big national orgs to sign on—the National Union of Healthcare Workers and Food and Water Watch—but that was out of a total of ~50 that I contacted or met with. Building a big AI tent that still centers x-risk remains a challenge; I’d be happy to hear anyone’s ideas about how we might approach this going forward.)
• Being too flimsy about the org structure early on. I was trying to be inclusive of many students who were eager to be involved (our first members were a student EA group and another AI safety group that has since had to step away from the coalition). This resulted in wishy-washy messaging (focused on too many issues rather than centering x-risk) and hesitant decision making. As the founder of the org, I should have been open to input but more willing to just set the strategy and assign roles.
• Focusing the coalition-building on orgs might have been a mistake. Even when organizations agree with our mission, it seems that it’s often difficult for them to join for various reasons. I think a coalition of a couple hundred individuals rather than a few orgs would have been easy to build in the same amount of time and probably would carry a similar amount of weight. For someone who doesn’t think they have the necessary connections to get a few orgs to join, this might be a better route. Tabling for a couple hours over a few weekends in busy public places would probably be enough to get a significant number of individual members. 

Skills Needed

I believe anyone can do this but a few skills are relevant:

• A strong layperson’s understanding of AI and awareness of the AI landscape.
• Decent communication skills. While being a strong communicator is helpful for coalition building, in work on other cause areas I’ve seen even below-average communicators have success with their legislators because legislative offices value constituent input. Don't let communication skills be the thing that stops you from doing this.
• Basic organizational skills to keep track of meetings, members, etc. 

Conclusion

It’s likely that loss of control incidents will keep happening. Having relationships with Congressional offices established in advance seems like it can be a good tool to steer those incidents towards meaningful action. Although there hasn’t been a tangible legislative outcome from my work with PAISC yet (no law passed or public statement made), my ongoing relationship with a Congressional office itself is a strong signal that this effort is needed and will bear fruit in the future.

Local coalitions won’t be a replacement for policy-focused orgs, but I believe they do fill a valuable niche that lays the groundwork for policy experts to do their job. It’s easy to get caught up in the EA bubble and think that most people already know everything that’s going on with AI: the capabilities progress, the risks, the proposed coordination mechanisms. After running this org for a few months I can tell you: they do not. Getting the ear of your local legislator in this way can funnel the right information to them and give them political clearance to speak out on this issue. One of them may even turn out to be an AI safety champion like Senator Sanders.

I hope some of you will start a local coalition like PAISC. Please reach out if you do—I’d be happy to hear about your experience and offer any help I can.