I’m writing this because many people are aware of the lump of labor fallacy and correctly reject it. But there are a number of scenarios around massive job reductions in AI that don’t rely on “we will simply meet fixed demand”, and I think it’s worth taking them seriously, and collecting them in one place. The cases below are from a world with plenty of demand for goods and services, but dramatically lowered effective pay relative to the present, for a meaningful chunk of the workforce. Lowered value can mean people getting fired, but it can also mean wages that can’t afford food and shelter, or just less dignity and fewer little luxuries.

1. AI can be a superior user of limited complements

How productive is a farmer with no land, no tractor, and no seeds? Not very. What stops AI models from being more effective users of land, tractors, and seeds than the best human? Nothing. The same applies to a manager of inventory, or a salesperson responsible for moving a given amount of product.

Capital is particularly harsh here because investors expect capital to have returns, and try to maximize those returns from the available options. Without active policy intervention, if AI continues to get better, human operational control over capital is likely to shrink. Assuming models are very law-abiding, humans can specialize in crime, and whatever niches we've made it illegal for models to fill.

Returning to the service sector, another type of limited complement is human time. If I am watching a movie, that is, implicitly, a decision that this is the best use of my time. It is not possible for most people to watch two movies, well, at the same time. Furthermore, the best human director and actors aren’t competing against the best movie an AI can make. They’re competing against the best movie an AI model can make for me.

Unfortunately, I currently expect that completely customizable and targetable media will beat high-quality work for most of the people all the time: utterly transparent slop is already growing in popularity1, and there’s a lot of room for improvement in the models. That doesn’t entirely eliminate jobs for human artists, but it puts them all in the position of an orchestral company or a dance troupe, performing for a few patrons and a primarily elite crowd.

Human labor that is restricted to an absence of scarce complements, whether human time or capital, leaves only tasks that are labor-intensive but capital-light. That’s a very slim set of jobs.

2. AI can improve faster than you can retrain

A common refrain in certain circles about AI-driven job displacement or loss is that we will just need to retrain the workers. Trucking is no longer viable? Let’s help people become home healthcare aides to the elderly (culturally difficult for white American men, particularly the sort most inclined to become truckers) or construction workers to build the datacenters! Oh, the datacenter construction process was 90% automated before the training program finished spinning up? Now we have two problems. This is an inherent fact of AI acquiring skills faster than humans do, and will persist so long as AI is both driving some humans out of jobs (probably already true on some margins) and improving faster than humans can (which is currently true and may continue for a while).

3. AI can monitor humans for free

Why are some people paid well, and others poorly? There are a bunch of factors, many of which I’m going to skip, but one of the less obvious ones is that there are many jobs where it’s very difficult to tell if someone is trying their best, or putting in the bare minimum to not be fired this quarter. In that regime, companies will pay very well so that employees think that having the job is much much better than being unemployed, even if they don’t like the work.

There’s another regime that workers can be in, aside from the loyalty regime. I think of it as the monitoring regime, which Amazon warehouses have perfected. Bathroom break takes too long? Penalized. Slightly slower than your maximum possible speed? Penalized. The roots of the approach date back over a century, but the key thing from an employer’s perspective is that replacing bought loyalty with monitoring can save a lot of money.

AI is going to be really good at rapidly going through an eight-hour screen capture of a white collar worker’s screen and identifying stretches where they were slacking off, setting up a doctor’s appointment on company time, or just not doing anything on screen while not in a meeting or an approved break.

The same job will pay less, punish little breaks more, enable more tyrannical bosses, and be more resistant to employee organizing.

4. There is infinite demand for human labor. There is no requirement that it pay enough to live on.

You may be familiar with the Law of Comparative Advantage. Even if you are better than me at every task imaginable, because you can’t do them all at the same time, we can both be better off via trading. It’s one of the most uplifting and inspiring laws of economics. However, if you add a constraint that I must consume so many calories and take up so much space, I may not produce enough value with my labor (particularly very low-capital labor) to survive. I have to be able to beat out shortform video for someone’s attention (at enough scale to support my life), or somehow make something valuable with the extremely minimal amount of capital I can make more efficient use of than AI.

But the situation is actually worse than that. If the price of certain items that were a large fraction of your budget (land to live on, land to grow food on and energy to grow it with) increases very quickly, because we discover new valuable uses for that land and energy, and your productivity grows at a slower rate (or falls, for the reasons discussed above), you will experience an effective pay cut.

What should we do about this?