Thank you very much for writing this.
Note: This was written quickly and is mostly experiential. I am US-facing and less confident about other systems of government, but the advice should apply to any representative democracy. I consider myself on the outer rings of both bay area AI and DC politics/elections, perhaps a tiny DC tilt.
On what to do: I firmly believe one of the highest-EV advocacy actions people can take is to call their national representative’s office. Look up your congressman, dial the number, and spend ~8 minutes on the phone trying to convince them to regulate AI for x-risk (I recommend using the OAI-HuggingFace hack to argue for increased CAISI funding if you don’t have an alternative agenda).
It’s a tangible action that takes a fraction of the time it took to read this post.
The backlash risk is small, as local offices often filter out true ‘crazies’ (think deep-end conspiracists) so the bar for harmful outreach via this mechanism is quite low.
It’s a second independent vector for advocacy. You count as a private citizen, and barring extreme cases no one will dismiss that because of a primary role in safety. You also don’t need to explain your day job.
This is literally the mechanism representative democracy was designed to operate through. You have a representative with an electoral duty to faithfully represent their constituents’ beliefs. Use them.
The upside is high. Even with a single call, you could plausibly filter the AI conversation directly into your representative’s ecosystem (“guess what that constituent just spoke about...”)
The collective upside is very high. If every member of the AI safety community took this one low-friction low-cost outreach every representative would have multiple independent touchpoints. Ripped napkin math: BlueDot safety slack has ~10k members. If each called and 1-in-4 are American that means 2.5k contacts for 435 representatives, or 6 average independent touchpoints per rep (likely clustered, which is fine).
Note this is a floor, nothing stops you from calling each month.
The president cannot write law that persists past his presidency. It takes 218 congressmen, 51 senators, and one president, or 290 congressmen and 67 senators to override the presidential veto.
(I’m saying America’s national representatives belong in your ~1000 belief funnel.)
A higher-EV action involves showing up to your national representative’s office. For a constituent, they will have someone actually meet with you and listen, it’s the definition of their job, and if you live in a remote area there’s a nontrivial chance you get the rep themself.
Yes this is higher cost, but that cost is directly converted into a credibility signal to the rep/office. Showing up signals that it matters enough to show up.
Busier reps are harder to get but conversely higher-leverage once moved.
America-specific estimate: | in-person attendance in office (or town hall) >> call > email; house representative > senator > state rep |
I’m really not kidding. I’d ask if anyone can formulate a strong argument against it, but the opportunity cost of writing that argument likely covers the cost of the call itself. (You can argue it’s not likely to move the needle, and this is true at the unit-level, but if it takes more than five minutes to write or evaluate that argument why not call anyway?)
Other thoughts:
I think AGI/RSI/loss of control is not an optimal target for level 1, it’s just a very high bar for the average politico. I’d suggest a floor of (AI-safety coded) autonomous CBRN uplift for adversarial actors, or (DC coded) the risk that a rogue nation-state or terror cell uses AI to develop bio/nuclear weapons.
Nuclear conflict is plausibly the single greatest fear of today’s legislators. It has a 60-year history of doom and is already accepted as existential. ‘AI could lead to nuclear proliferation’ is a much much easier sell than ‘AI could escape the lab and end humanity.’
Similar for bio risk, though references to covid or the black plague will earn more credibility than mirror life. Anecdotally: mirror life codes as sci-fi; ‘more lethal covid’ maps onto lived experience and credible lab-leak.
Indexing too much on convincing people of rogue AI risks overfitting, for lack of a better word. The goal is political will --> robust law, and you don’t need this specific scenario to vote in favor of a strong law. If you can break people out of the sci-fi frame, then great! But it’s a lot easier to sidestep it (and I’d argue an easier onramp).
I can’t name a single person who has sold a political operator on (superintelligence + rsi + loss of control = xrisk), but I have convinced someone on (terror cell + unsafe capable AI = probably massive virus and millions/billions dead). The response is not “well, technically that’s global catastrophic risk not xrisk so I’m not sure it’s worth prioritizing” the response is “oh fuck.”
I can see the outline of an objection that a action built on less-important risks will not address enough. With low confidence I think the policy wonks who write the crisk/xrisk legislation are likely to draw on preexisting resources and include all the risk vectors as long as leaders prioritize the general topic, though I’m very fuzzy on this and open to a rebuttal.
I strongly recommend against paying people to do outreach this way. It’s misaligned with representative democracy, has a much higher blowback chance (“EAs are trying to buy AI regulatory capture!”), and overall less effective (paid advocates who don’t believe usually fail the vibe check political-world people use as a screener).
I’m not against advocacy research per se, but I worry people will fall into a trap of using research to justify not talking to people themselves. My strong prior is that the marginal unit of effort spent on persuading people is more valuable than the marginal unit spent on figuring out how to persuade people. Just call your rep.
I was also at this event and can confirm everything Ryan said! I had a very similar experience and was a little astonished by how with it the attendees were. The event was under Chatham house rules but I will note that the ’we’ve all read LessWrong” speaker is very much a DC native not some random EA person who just happened to be there.
I will also recommend in their workshops or applying for their fellowship (closes Aug 30th, begins Jan) if you have any interest in policy, it can be as short as a six month DC ‘tour of duty’ and I think there’s a lot of benefit in growing the SF/DC style connective tissue.
Otherwise my big takeaway was about the importance of trust and relationships. If you want to enact any sort of law, the message they gave us is that your experience/status/title all matter significantly less than whether the people who would write/vote on it trust you and like you and will let you work on it (or better, come ask you about it). That can come from your pedigree but it’s mostly about relationships. I think this is why DC cares so much about coffee chats and networking; credentials and artifacts are swapped for referrals and relationships as trusted filtering mechanisms.