I read a paper this week called The AI Layoff Trap. It has an interesting thesis.

If you look at what is happening with automation and laying people off, and take it to the ultimate conclusion, no one will have a job. If that happens no one can afford to pay for a company’s goods or services. So, logically, it doesn’t make sense for a company to automate all the jobs and lay everyone off. If they do that, they have no customers.

The paper argues that even if it doesn’t make sense, businesses have no choice but to do it. Automate and lay everyone off - and have no customers.

When a company automates work and replaces people with AI, it keeps all of the savings. The laid off people have less money to spend, but not all companies are moving at the same speed, so there are still employed people who can spend money. The loss of demand from the people laid off is spread over all the companies providing the services.

This leads to an arms race. Everyone agrees that mass layoffs will eventually hurt them, but nobody wants to be the one company that holds off whilst its competitors get cheaper.

The paper suggests a Pigouvian tax on automation - charge companies for the cost they’re pushing onto everyone else when they swap human labour for machines.

I think there’s a big problem with that - how do you measure automation?

Say a company employs 10,000 people one year and 8,000 the next.

How many of those 2,000 jobs went because of AI?

The company will tell you it was restructuring. Or outsourcing. Or natural attrition. Or a soft market, or a merger, or just a bit of belated efficiency improvement - the old “we overhired and now we’re just correcting”.

And things are never simple in business - nobody sits down and decides to make 2,000 people redundant for one clean, well-documented reason. There’s a whole bunch of horse trading and carve outs that makes it almost impossible to know the real reason.

Working out which specific jobs were “lost to AI” turns into an accounting exercise, then a lobbying exercise, and finally a legal one. The elegant tax starts to look impossible to actually collect.

Here’s the alternative: don’t tax the job losses, tax the AI.

A very small levy on inference - tokens consumed, or more sensibly the compute needed to produce them - collected from the model providers and the big compute platforms. Nobody has to audit anybody’s HR decisions.

It’s a blunt instrument. A company using AI to make its existing staff more productive pays the tax without laying off a single person. So do researchers, startups, and people doing things nobody has thought of yet.

You lose precision and you gain something you can actually measure.

The rate doesn’t need to be big either. The point isn’t to make AI expensive enough to stop automation happening - the productivity is real, and trying to hold it back would be both futile and a bit daft. Think of it as a small levy on something that’s quietly becoming a factor of production alongside land, labour and capital.

The paper’s model has a nice property: the problem shrinks when displaced workers get more of their income back. Retraining, wage insurance, actually finding people new jobs - all of it reduces the hole in demand that automation creates.

So the loop looks like this:

AI usage → small compute tax → worker transition fund → more income replacement → less demand destruction

The tax never has to work out which API call replaced which person. It just skims a little off the top of the activity that’s driving the transition and puts it towards helping people through it.

What are the problems? And there are plenty of them…

Open-source models and self-hosted inference are awkward to collect from. Offshore compute is an obvious escape hatch. And deciding what actually counts as taxable AI compute would keep a room full of lawyers busy for a good while.

None of that is trivial. But it feels more tractable than asking a tax inspector to rule on whether a particular redundancy was caused by AI.

A token tax would be theoretically wrong in all sorts of ways. But a slightly wrong tax on something you can measure is probably better than a theoretically perfect tax on something you can’t.

What do you think? Let me know in the comments.