What if AI Safety employees unionised?
American bald eagle caws angrily in the distance
Whoa whoa whoa, just hear me out. Unions aren’t usually a good answer for free-market loving libertarians, but one particular AI safety problem is awfully union-shaped:
Repeatedly, companies have started out being pro-AI safety and talked the talk about how they’d take precautions around building advanced AI systems. Repeatedly, this line was used to placate & hire very talented researchers who cared about AI safety. Repeatedly, these companies did not in fact take precautions around building advanced AI systems.
Imagine you’re a researcher who’s concerned about the risks posed by advanced AI (maybe you don’t have to imagine). Imagine further that you join a company in large part due to their strong safety commitments. Wary of value drift, you write down your red lines: “If my beloved company starts racing to AGI, I’ll quit”. You even tell your new boss about your red lines. He nods seriously.
A few months go by, and alas, your company reneges their AI safety commitments. This leaves you in a bit of a pickle:
You could threaten to quit, but they’ve publicly turned away from AI safety and you alone are unlikely to change their minds. You could say you’ll quit next time they cross your line in the sand, but that threat is less convincing the second time round.
You think to yourself:
If only there was some sort of conglomeration-like structure that could represent all AI safety employees, allowing them to raise their complaints and actually be heard. Hmmm… Such a conglomeration could credibly threaten or even veto moves by the company that would be decidedly unsafe. This would be great!
If not a union, why union shaped?
Unions are designed to allow employees who face a mass-action problem to combine their bargaining power and steer their company in a better direction. Usually this is used to get benefits like greater wages, more time off, more work, or less gruelling work. But possibly, a union (or something shaped like a union) is an existing mechanism that can be applied to let researchers credibly threaten their employer if the employer seeks to take actions that increase the risk posed by advanced AI.
Unfortunately, all is not as clear as I made it out to be. Unions don’t quite work as nicely as I described:
Business decisions don’t have to go through the Union
Unions can’t negotiate around business strategy nor corporate governance, only things like wages, hours, and “other terms and conditions of employment” (NLRA§8(a)(5)). Plausibly “other terms of employment” could include safety-relevant factors, but this seems a bit of a push.
Manager/supervisor-type jobs can’t be in a union
NLRA §2(11) says (roughly) that supervisors can’t be in a union, and this likely includes anyone who leads a team or has authority over other employees in some way. So leading a research team would preclude you from joining an AI safety union (although you likely have more negotiating power if you’re a team lead).
Visas, strikes, and getting fired
Many AI researchers are on visas, and cannot risk being fired[1]. Even the researches who aren’t on visas likely have a significant amount of wealth tied up in unvested stock options, and won’t quit for a small infraction. The employees risk a lot by threatening to quit, so any credible threat of quitting has to be matched by a serious issue with the internal safety policies of the company.
A more realistic alternative to a union
Possibly, a union isn’t actually required. AI safety employees could (in theory) agree to resign if certain red-lines were crossed, and could make this agreement public for their employer to see. If this agreement was considered a credible threat and if many employees had signed the same agreement (as opposed to hundreds of subtly-different agreements), then an employer might take the contents of those agreements into account when considering business decisions that raise safety concerns.
Such a contract would need the following:
the terms of the contract must be public for the employer to see, such that the threat of resignation is predictable and avoidable.
the number of employees who’ve signed the contract must be known to the employer. Note that the identities of the employees aren’t necessarily important:
entry-level employees (or employees on a visa) could anonymously sign the contract, liaising with a third party to share the count of employees who’ve signed the contract with the employer
However, high-profile employees could lend credence to a contract by publicly signing it. Anonymity should be the default, it shouldn’t be required.
the contract should come into force only if some significant fraction of the employees have signed it, taking no effect otherwise. This removes the first-mover problem, allowing employees to sign the contract without worrying that they’ll be the only ones who quit.
Trigger conditions should be serious. Resigning will come with significant risks due to unvested equity and visas that are tied to the employer, and the reasons for resigning should match the risks.
Some very significant cost to the employee if they do not quit. I’m uncertain of what this could be, given that remaining with the company is a much better option financially.
Exactly how you phrase the red-lines is critical
The hard part, of course, is defining the red lines that triggers mass-resignation. I’m unsure what this would look like, given the diversity of opinions around AI risks. Ideally we’d find some red-lines that are serious enough that they indicate a meaningful risk, but not so serious that by the time they’re triggered it’s too late and mass-resignation won’t have any effect.
Note that employees would be giving up a lot by threatening to resign. Nobody’s going to give up millions of dollars in unvested equity just because Claude kinda blackmailed someone in an evaluation. If the red-lines are not sufficiently serious, nobody’s going to believe the threat.
Practicalities around signing the contract
A third party could adjudicate the signing, allowing employees to sign the contract confidentially, while still reporting the number of signatories to the employer[2]. The third party could be charged with deciding if the red-lines have been crossed, although this potentially just passes the buck instead of resolving the issue.
Having a large number of signatories could be a positive thing for safety-focussed companies, since it’s a costly signal that their employees actually care about AI safety. On the other hand, if such a contract were common knowledge and your “AI safety” company had no employees who had signed it, this would be a good indicator that the company doesn’t actually care about the risks due to AI.
Many problems circle around the issue of companies credibly signalling their commitments to AI safety. There are many dual-use research agendas such as mechanistic interpretability which a company might pursue in the name of AI safety, but later the approach changes when the company becomes interested in seeking a profit. An expensive mechanism which allowed companies and their employees to signal their commitment to AI safety would resolve many uncertainties around company agendas and charters.
Visas & employment protection
Note that non-profits could help by offering to employ people who are at risk of losing their visa and who resign due to their employer reneging their safety commitments. The annual H1B visa cap does not apply to non-profits, so a non-profit could sponsor employees as soon as they quit. This would reduce the downside to the employee for quitting. Grant makers like Coefficient Giving could aid here, by offering to fund the legal bills and sponsorship for H-1B & O-1A visas.
However, it could be catastrophic for the non-profit: they’d be committing to hiring a significant number of highly-skilled workers all at once, since this would only happen if a frontier AI company crossed the red lines and caused a large number of employees to quit.
It’s also illegal to pay someone without genuine employment. They need to be actually doing work, not just sitting around. I don’t think it’d be impossible to create genuine employment for dozens of AI safety researchers if they suddenly become unemployed. But pre-emptively planning to spin up an organisation that conveniently hires whoever quit their job is the sort of thing that Uncle Sam might not look kindly on. I assume there’s a way around this without doing shady things, but I lack the legalese to know what this looks like.
The employer’s best response
If a large fraction of your employees had signed a contract like one described above, your best move as safety-ignoring employer would be to specification-game the red-lines, finding as many caveats that let you continue to improve your AI without technically crossing the red-lines.
I think a contract as I’ve described above could provide useful signal about which companies are willing to take costly actions to build safer AI systems, and which are not. I suspect that what I’ve described is too naive in ways that I can’t see. I think that some variant of what I’ve described could work and reduce the risks posed by advanced AIs.
- ^
They are very unlikely to strike since NLRB v. Mackay Radio established that economic strikers can be “permanently replaced” — a crazy piece of legalese that means your employer can hire permanent replacements to do the job while you’re on strike. The job you had is now gone, which leaves you unemployed. Technically you get put on a “priority list” to be re-hired. I’m doubtful about whether you’d actually be re-hired. There is a slight caveat: strikes due to unfair labour practices are different, and do require the employer to give you your job back after the strike ends.
- ^
There’s some interesting encrypted-information protocols (see Callisto) that originated as a way to let sexual assault be reported anonymously such that a lawyer is only contacted if more than k people report the same person. This approximately removes the need for a third party, although still requires someone who acts on the information once more than k people have reported it.
Interesting—thanks for writing this up. This updates me against thinking that organizing lab employees is that valuable of an intervention. Unvested stock options makes the incentives so bad… I struggle to imagine a scenario where you have a large fraction of lab employees willing to give that up.
Are there are other things lab employees could do besides threatening to resign?
I think the question of whether lab employees would resign is a bit subtle, but overall doesn’t look great. Employees do resign for ethical reasons(!) but unfortunately the act of resigning also removes all their bargaining power with their previous employer, making resignation a nuclear bomb sort of strategy.
Other things the lab employees could do? I’m not sure. The US doesn’t make it easy for an individual to sway the business decisions of the company, and that’s basically what we’d lab employees to be able to do.
There are softer options (talking to the higher-ups, voicing dissatisfaction, advocating for better stances) but these don’t seem to really move the needle
completely unrelated to the article, but surely libertarians are against laws which protect unions, and not the practice of free assembly, right?
I think it is generally bad for suppliers to collude to fix prices, whether those suppliers are companies or laborers.
how do the proverbial fishers avoid overfishing, except by colluding to limit supply?
(i’m not trying to be obtuse here—genuinely curious whether there’s a neat principle by which these can be distinguished!)
(I don’t have a stake in the libertarian fight, but, I recommend reading the intro and 1-2 chapters of Elinor Ostrom’s Governing the Commons, which argues “there are tons of special snowflake ways of handling overfishing-etc without exactly establishing a central authority or privatizing the commons.”
on a skim, the proposal is for the fishers to voluntarily hire an enforcer / voluntarily agree to certain contracts regarding the deployment of fishing boats. i agree that this is neither a central authority, nor privatization. but isn’t it exactly a union?
the suppliers of a good/labor voluntarily agree to certain provisions on the production of said good/labor for their own long-term benefit.
(i really appreciate the resource, and will give it a closer look!)
Well the point is there are a lot of different examples that have different little fiddly details.
could you point to one that’s not essentially “agree to a reasonable division in an ultimatum-like game”, or help me understand how such an agreement is not essentially what a labor union is doing?
Nope, I did not actually read it that thoroughly.
But, I think you’re sort of prematurely/overly lumping things together (see: lumper/splitter), which I think is erasing detail that is worth actually thinking through.
The point is the people in all the different situations did not start by asking “do we want to handle this with labor unions or government centralization or privatization”, they started by working through the details of the individual cases, and maybe the stuff ends up sharing some structures or could be ontologized a particular way, but the details of what the people-on-the-ground want/need and the realpolitik matter a bunch.
i guess this is what i’m trying to understand.
as a meta point, i get the impression that you think i’m being obtuse for rhetorical reasons. but i promise you: i’m just obtuse!
like, based on this comment, it seems like you think i’m going “today i will lump these two obviously different things together, he he he” (i guess because i want to lose some arguments on the internet :) ). but from my perspective, i can’t tell the difference! everyone’s telling me that there’s a clear distinction between labor unions and fishery contracts, but no one is telling me what that difference is. i’m confused!
Okay lol, fairnuff.
The thing I vaguely remember from the book’s early chapter was, a fishing town where they rotated slots you get in the lake each season. I can’t remember if that was a thing the town elders/government decided sort of from on-high, or that the fishermen negotiated together and organically turned into a norm, or what.
But, the point was “rotate slots in the lake” is a mechanism that is neither exactly private property, nor setting a limit on number-of-fish. This doesn’t have much to do with labor unions.
I believe there were 1-2 other mechanical details in that chapter but I don’t remember, that were similarly “oh, huh, I guess you could do it that way.”
(The implication here not being that AI safety unions should use rotating slots, which is nonsensical in this case – I was mostly replying narrowly to the “how do the proverbial fisherman avoid overfishing without limiting supply?”. With a bit of a broader implication for “mechanism-space is deep and wide and local-details-space is deep and wide”, which might or might not be relevant to the OP idk)
That would be a negative externality, which is one of the situations where “generally” does not apply. The standard economic solution is to internalize externalities, e.g. the government imposes a tax on fish, such that overfishing is no longer profitable.
i feel like the dockworkers, or whoever, will argue that unless they bargain collectively, they’ll end up selling their labor for less than the cost of living. this will have the negative externality that there won’t be any dockworkers (or at least no one will apprentice to become a dockworker).
note that this is empirically verified, for example with air traffic controllers: they are legally prohibited from striking, and now nobody wants to be an ATC. we pay this cost in lives!
That’s not an externality, it’s an...internality? Is that a word? The people who would be harmed are the people who are demanding dockworker labor. If this is a big enough problem for them, then they’ll be willing to pay more.
can’t you say the same about the fishery? the people who would be harmed are the people who are demanding fish from the lake. if it’s a big enough problem for them, then they’ll be willing to get up earlier and fish for longer.
the collective bargaining is the way that the future harm is brought into the present.
since the harm is in the future, it’s at remove. someone else’s problem. an externality!
the situation would be different if it was easy to scale up and down the number of laborers just by inputting more money. but it’s not—it takes investment of time through training. for example, air traffic controllers require six years of study.
it occurs to me that we might not be in disagreement at all. could you clarify that the claim this comment is making is that “free-market loving libertarians” specifically are against collusion, including collective bargaining as unions?
if this is rather a personal political opinion, then i will say that you may very well be right! i simply do not understand the issue enough to have an opinion either way. my initial claim was narrowly about what a principled libertarian would say, and not at all intended on behalf of any specific policy.
The grandparent comment is stating my opinion. I can’t speak for all libertarians. There are some libertarians who say all government intervention is bad. My position (which I think is a popular position on LW, and popular among economists) is that governments should support free markets, which de facto looks like libertarianism most of the time because most government intervention makes markets worse, not better. But there are some circumstances where government intervention is justified, like breaking up monopolies and regulating externalities.
thanks. that all seems reasonable.
regarding labor unions, do you think the government should regulate their formation, since they involve price-fixing collusion?
I don’t have a strong opinion on whether the econ 101 arguments are the strongest consideration, but I weakly believe the government should ban labor unions in the same way it bans price-fixing collusion in general.
i see. so if the union obtains monopoly pricing power (according to some definition), it should be broken up then?