Maybe job displacement is an extreme risk

Discussion of AI and jobs doesn't focus enough on the risk of total human disempowerment

“You have said...’development of superhuman intelligence is probably the greatest threat to the continued existence of humanity.’ You may have had in mind the effect on jobs.”

- Senator Blumenthal, speaking to Sam Altman during a Senate hearing on AI in July 2023

The above quote was, quite reasonably, mocked all over Twitter (it stuck in my mind enough for me to find it 2 years later!) Sam Altman was of course talking about the risk of AI killing all of humanity, not the ‘mere’ threat of job displacement.

Job displacement and human extinction certainly sound like very different types of threat. But what if they’re not as far apart as they might seem? At the extreme, if AI replaces almost all human work, we might find ourselves in a world where we humans are so irrelevant and so powerless that we have very little control over the resources we need to survive.

I started thinking about this recently when reading a great recent(ish) paper called “Gradual Disempowerment” by Kulveit, Douglas, et al. They talk about the possibility that humanity might end up “losing control” of the future to AI systems not via some sudden, decisive power grab, but by gradually handing over more and more crucial parts of society to AI. This is a kind of risk scenario I’ve been concerned about for a while, and the authors do a better job explaining it than I’ve ever managed to. I highly recommend reading the full paper.

The paper talks through how we might realistically end up handing over control of different domains to AI and what this could mean for humans. The part that particularly stood out to me was the section on how we might lose control of the economy, which argues that:

• In future, we may see AI systems capable of performing most cognitive tasks more efficiently than humans;

• This could lead to a kind of job displacement that is qualitatively different from anything we have seen historically: “rather than just shifting the type of work humans do, AI could potentially reduce the overall economic role of human labor”;

• If humans play an increasingly small role in the economy, that massively reduces the power they have to influence it. We might then find ourselves in a world where AI systems “command a growing share of economic resources and make an increasing proportion of economic decisions”

• In such a world, “the economy might become so optimized for AI-centric activities that it fails to maintain infrastructure and supply chains which are critical for human survival”.

This story of course relies on some assumptions: that we will get to the point where AI systems can do most cognitive tasks better than humans; that a large majority of economically valuable work will in fact be automated; that humans won’t be able to retain some high-level control over AI systems in the economy, to mention a few. I think these assumptions are non-trivial, debatable, but not at all unreasonable or even particularly “sci fi” given current trends (if you disagree with me, I’m very interested in hearing why!)

While there’s increasing discussion and political interest in the problem of job displacement, I don’t hear many people talking about it in this way - as something that could be catastrophic or even existential.

Anton Leicht writes about the lack of good policy proposals for dealing with job displacement today, despite it being a very politically tractable area of AI policy. I wonder whether this is partly because many of the people who are seriously concerned about extreme risks from AI don’t see job displacement as something that’s in their wheelhouse or even particularly neglected. There's certainly a lot of general chat about AI and jobs, but I think the question of how to ensure job displacement doesn't end up being a route to total human disempowerment seems highly neglected.

If we frame the ultimate risk as being about human disempowerment, I think the policy solutions probably look quite different to many of those being proposed today. This obviously needs a lot more thought than I can give in a short blog post, but some quick thoughts:

• We should think harder about whether there are particular categories of job that shouldn’t be automated from the perspective of humanity as a whole retaining important types of power in the world (e.g. parts of politics, the economy, infrastructure). This makes less sense if you’re concerned about the impact on individuals or groups of not having jobs, which is more how the problem is often framed - why would some matter more than others? It might also plausibly be easier to build political will to ban certain kinds of information if people are really convinced that what’s at stake is humanity’s control over the world.

• We might want ideas for alternative economic/political/social models that don’t tie power as closely to work as our current capitalist system does. I’m not an economist/political scientist so I don’t even really know how to think about this, ... but when we’re talking about a future with advanced AI it’s not at all crazy to consider we might need to organise society in very different ways to today. Currently discussed proposals like UBI focus on how to ensure people without jobs have the money they need to live within the current system, but that’s very different from asking how to ensure people have power in a society where they don’t have jobs.

• There might be some interesting coalitions / bridges to build here between groups who are concerned about certain types of people becoming more and more disempowered and marginalised in society (e.g. the technically illiterate, disabled or mentally unwell) and those concerned about humanity as a whole losing power to machines - at the end of the day, the question of what kind of power matters to live a good life, and how to divorce that from work as we know it, is probably the same for both.

The paper on gradual disempowerment I mentioned above also has some great suggestions for reducing the risk of human disempowerment in general, including ways to measure human disempowerment - for example by measuring AI share of GDP. As with all risks, being better positioned to assess how (good or) bad things are would be a very good place to start.