A superintelligence bill was just introduced—now is the time to act

[dh]

Sen. Bernie Sanders and Rep. Greg Casar have drafted a bill to ban superintelligence and pause AI development until a regulatory body is in place.

It should not be understated how important this is for AI safety:

• This is the first major AI legislation proposed in the US, and therefore, the first that would directly impact frontier labs,

• It is a piece of legislation that explicitly takes into account x-risk AI safety considerations—superintelligence, an AI pause, biorisk,

• Bernie Sanders is a mainstream politician with a large following, giving an opportunity to significantly shift the Overton window.

This specific moment is particularly opportune, occurring at an intersection of all three streams of Kingdon’s multiple streams framework for policy change:

• Problem: the OpenAI/​HuggingFace hack (and the drop in Astra monitorability)

• Policy: the given bill

• Politics: AI popularity is at an all time low ^[1]

If ever there was a time for AI advocacy, it is now.

Why we should act on this:

• This bill failing to gain popularity chills out future AI discourse. When people hear superintelligence, they will think “That was that thing insane Bernie was talking about right? What a crazy guy!” (this is already happening). This is probably the first interaction for a lot of people to AI safety, and first impressions are everything. We can’t allow this to become a politically toxic.

• Very few people know about AI safety as a field, even amongst AI researchers. In their mind, the experts are computer science professors, or worse, frontier lab researchers. They should not be allowed to set the agenda—if discussion gets bogged down in “LLM’s are token generators that can’t think real thoughts”, very little progress will be made.

• There are good objections to be made about the bill. If you disagree:

  • Talk about it! Contact Sanders’ office, tell them what you want fixed, don’t just put it in a comment that will never be seen by a politician. Contribute to the discourse to make this better.

  • This doesn’t mean you can’t support the principle of the bill. If you aren’t on the news saying the idea of the bill is good, then someone else is saying the entire thing is garbage.

  • This policy only improves with discourse, and public discourse can only happen if the bill is taken seriously in the first place, which requires outreach to legitimise the discussion.

  • Most bills are completely different from the original proposal, so these disagreements can be ironed out later.

• We should reward politicians who suggest good policies. Yes, this might be some trivial benefit, but Bernie has taken us seriously, and should get some props for that. If the only thing that we do is nitpick and infight, there is little political reason for him, or anyone else, to push for this any further.

• A lot of people agree with this! It’s not just you. It’s hard to take the first step, but you won’t be sticking your neck that far out, and people will support you.

What should we do:

Research organisations (METR, Redwood, MIRI, GovAI, CAIS, FLI, AI Futures, etc)

• Reach out to Sanders’ office and offer your support for the bill.

  • Having experts’ support this is incredibly important for success. Having the majority of experts is even better. These are the people who are asked to testify in congress, who are contacted by other politicians.

  • If you have complaints, raise them, but tell him that you support the principle.

  • Even if the average person hasn’t heard of you, people tend to defer to experts.

  • This would probably matter for some people within AI, and whether or not they see this as a legitimate proposal or some normie bullshit.

• Tell everyone that you support this (and if you disagree with some components, endorse with whatever clarifications you want). Write a public statement, send it news organisations, post it on your website. Sign an open letter, whatever. Make yourself known.

  • This is an excellent opportunity to get your name out there. Yes, doing PR is a bit demeaning, but no-one will listen to your research unless they know you exist. Getting your name out there gets you invitations, it gets you donations, it gets you in the room.

    • If you have to, hire an intern for a small stipend to do all the outreach.

  • When reaching out to people, flex your achievements. Link and explain some good research or cool results that you have (e.g. scary misalignment results). Show them why they should listen to you, and why they should be afraid of AI.

  • Make it clear you are independent of labs. You aren’t part of “Big Tech”, you’re trying to fight against them. People will distrust you the more they think you are just “building AI hype”.

  • Send it to many congresspeople, to mayors of cities, to the heads of the parties. In politics, everyone is fair game to fight for your issues.

• Note, I have no idea how much of this, if any, has been done already behind the scenes.

Individual researchers:

• Contact your congressperson.

  • This is a bill they will have an opportunity to voice an opinion on. Make it clear that you support it.

  • Explain why you should be listened to. If you are an expert, specify this. Offer to have a chat with them about why it’s important. If you don’t want to do that, then contact them and point them in the direction of someone who does (I’m sure ControlAI or other advocacy organisations could find someone to brief them)

  • Contact Sanders’ office too—they will need a LOT of people to defend their stance, and it is much harder for them to find you than for you to find them.

• Contact the news—no matter what country you are in.

  • The largest political movements gain traction internationally, even if they only affect the US.

  • Tell them you’re an expert, that you think this is a great policy (or, the idea of the policy is great). Offer to be interviewed, or point them in the direction of someone who can be interviewed.

• Contact content creators

  • I’m sure most of us follow some set of STEM content creators. Tell them about this. Offer to explain the issue to them, or link some documents summarising the problem (it would be better if they were succinct, so they don’t need to commit a lot of time upfront).

  • See if they will spread the word about this. Policy is not just a discussion for politicians, it relies on the opinions of the public.

• Talk to people around you

  • Talk to family and friends, coworkers. Yes, it’s unlikely this will change much, but it helps you practice having these discussions for when they do matter.

• If you work at a lab, express your support

  • This is a tangible manifestation of the “pacing the frontier” letter. If you are safety minded, now is the time to show it.

    • This is also the time for your labs to show it. If they claim to care about safety, then they should be happy with this proposal. Even if it isn’t perfect, they should be working with politicians to make it reflect what they want, not outright denying it.

  • Even if you report anonymously, being an insider who agrees is very powerful for the policy narrative.

Grantmakers:

• Share contact lists

  • Make a list of who is in your network, and send it to politicians or news organisations. AI safety is diffuse, so make it easy for them to find an expert.

    • Specifically, different types of experts (research scientists, university professor, policy maker, congressional staffer, advocate, etc)

• Donate to advocacy

  • Pay content creators to talk about this, run articles in newspapers, help shape the narrative

Note on polarisation: The White House has recently demonstrated their desire for a light hand on AI. Given this bill was proposed by to Democrats, and it involves regulation of a large firm, it is likely that this falls on party lines. We cannot let this happen. We must make it clear this is a non-partisan issue. Both sides are concerned about AI, so this does not have to be inevitable. Explicitly reach out to Republicans. Make sure they know you support this.

There are a lot of reasons this bill isn’t perfect, and I want to emphasise this is not what our finalised AI regulation should or will look like. However, if ever there was an opportunity to get people to care about AI before it becomes too dangerous, that time is now.

1. ^

   The reasons why this is the case don’t necessarily line up with the Problem and Policy stream, but that suggests we should attempt to bridge this gap through awareness

[MichaelDickens]

This is the first major AI legislation proposed in the US

I don’t think that’s true? Unless you have some specific definition of what “major AI legislation” means? The AI Kill Switch Act was introduced in July, which I would think would also count by your definition.

To be clear, I like the Sanders & Casar bill, I just don’t think this statement is true.