The lump of labour fallacy fallacy

Every time someone says the machines are coming for the jobs, there’s a guy in the back of the room who knows the term. Lump of labour fallacy. He’s read his economics and he’ll explain it to you slowly, the way you explain things to a child. Work isn’t a fixed pie, he says. Automate one job and the economy invents three more somewhere you couldn’t have predicted. It happened with the loom, it happened with the spreadsheet, it happened with the ATM – banks hired more tellers after the cash machine, not fewer, look it up. So relax. Every century the automation take the jobs, and every century we all somehow still have work.

He’s right. He’s been right for two hundred years.

That’s exactly what worries me.

But here’s the part that guy in the back almost never gets. The lump of labour fallacy is a real fallacy – people genuinely assume there’s a fixed amount of work in the world, and they’re genuinely wrong about it. But there’s a second fallacy, hiding inside the first one, and few actually see it right away. An economist did, though. Daniel Susskind, in A World Without Work, calls it the “lump of labour fallacy fallacy”, and once you see it you can’t unsee it.

It goes like this. The lump of labour fallacy is the mistake of thinking the amount of work is fixed. The lump of labour fallacy fallacy is the mistake of thinking that the new work that gets created has to be done by humans. The fallacy was never about whether more work gets created. Of course it does. The fallacy is the quiet assumption that we’re the ones who’ll be doing it. Strip that assumption out and the whole reassurance collapses, because there’s no economic law that says a freshly created job has your name on it instead of an agent’s.

Susskind reaches for horses to make the point, and it’s a brutal little metaphor. For thousands of years every new tool made horses more valuable. Better plows, better carts, better roads – more demand for horses, every single time. The lump of labour fallacy applied perfectly to horses, right up until the engine arrived. And when it did, the horses didn’t retrain into more leveraged, higher-judgment. high-accountability horse work. Today, the horses are simply unemployable. Suskind is not the only one saying it, either. Calum Chace makes the same argument from the AI side in The Economic Singularity – yes, new jobs will appear, but nothing guarantees humans get to keep them. The escape route that saved us every previous time was that machines took the muscle and left us the cognition. What happens the day the machines come for the cognition? Because – if you haven’t noticed – that’s precisely what machines are after this time around.

I gave a session at DynamicsMinds in Portorož this year, “The Future of Software Development,” and I’ll spare you the suspense. I’m not optimistic. I think within the next two to three years, agentic tooling gets good enough that most of what we currently call software engineering becomes something an agent just does. Not all of it. Most of it. How fast it lands depends on a hundred things none of us can predict. But land it will, sooner rather than later.

And now the part that makes the survivors uncomfortable, including me.

I’ve been reluctant to try ChatGPT. Today I got over that reluctance. Now I understand why I was reluctant.

The value of 90% of my skills just dropped to $0. The leverage for the remaining 10% went up 1000x. I need to recalibrate.

— Kent Beck 🌻 (@KentBeck) April 18, 2023

I hope I don’t need to introduce Kent Beck. I agree with him 100%. Not just “he has a point” kind of agreement. No. He described, precisely, what will happen (and probably did already, to an extent) to every one of us.

But April 2023 was the start of this trajectory, not the end of it. I mean, seriously, a lot has happened in three years so it feels weird explaining it, but here I go anyway. The 90% that went to zero between 2023 and 2025 – the syntax, the boilerplate, knowing where the semicolons go (I am drawing a caricature here, but you get the point) – that wasn’t the end of it in the sense “we got the new tools, now we get to use them for decades”. Nope. It was only the first step. Because that precious 10% everyone’s clinging to now, what is it? There’s nothing sacred about it. It’s just the part the tools couldn’t reach back in 2023. And tools reach further every quarter. So ask yourself honestly: is there anything about the 10% – the architecture, the product, the industry, the “experience” (I have rarely encountered genuine experience, something that nobody ever experienced before – and then didn’t documented in a book, blog, or an article), the taste, the knowing-what-to-build – that the same curve won’t eventually swallow? Not today. Not this year. Maybe not even 2027. But look at the trajectory and point to the segment that bends away from the work you’re so proud of. I can’t find it. Most of what I did for 30+ years is now being done by agents – and it’s awesome, because I don’t waste time on typing code anymore – I imagine things, describe them, plan them with an agent, build a spec/plan, build the agentic scaffolding that keeps my agent tightly constrained under specific guardrails, and I am overall more productive. But, as I said, I don’t think it stops here. A year ago there was no scaffolding. That was the 10% of the last year. But I think the 10% goes through its own 90/10 split, and then the survivor of that goes through another, and the splits don’t politely stop at a number that happens to feel safe to you and me.

You still think the new roles will stay ours? Fine. Where are all the prompt engineers?

Eighteen months ago that was the future. The role everyone was going to migrate into. Conference tracks, LinkedIn titles, paid courses, the whole circus. How long did it last? The agents write the prompts now. Better prompts than the prompt engineers wrote, at three in the morning, without asking for a coffee break. The shiny new job appeared, got automated, and vanished inside the same window we spent congratulating ourselves for inventing it. Now look at spec-driven development, the current great hope. Write the spec, let the agent build it, best of both worlds. It’s good – it really is. But go count what’s actually inside that loop and tell me how much of it is genuinely human. The spec gets drafted by an agent, refined by an agent, implemented and tested and reviewed by an agent. We keep finding the human seat at the table and noticing the chair got smaller while we weren’t looking.

So what’s actually left? Three faculties are being floated around these days:

• Judgment – deciding what’s worth building and what “good” even means
• Governance – setting the rules the agents run inside
• Accountability – being the name on the line when it breaks, because you can’t sue a model

Don’t get comfortable there either, because those three are very much layered, and the stronger the tools get, the more layers they expose and quietly shed. Today you exercise judgment over a function. Tomorrow over a module. Soon over a whole system, then a whole portfolio, and every step up the ladder means one person doing what used to take five. That’s not work multiplying. That’s work concentrating. And concentration, for everyone who isn’t standing at the top of it, looks exactly like the door closing.

Let’s take judgement, for example. Genuine new ideas, the thing no machine can supposedly do. Take a look at what happened to Erdős Problem #1196. It’s a 1968 conjecture from about primitive sets, that eluded the best of mathematics minds for sixty years. Jared Lichtman spent seven of them on it. Then in April 2026, GPT-5.4 Pro proved it, the result was formally verified and published. The model found the proof in about eighty minutes and wrote it up as a LaTeX paper in another thirty. The person who prompted it, Liam Price, was a 23-year-old with no advanced mathematics training. And here’s the part that should keep you up at night, because it kills the “it’s just fancy autocomplete” defense stone dead: the model reached for a Markov chain technique that human mathematicians had overlooked despite years of work on the problem. That’s not retrieval. That’s a new idea. And 1196 isn’t a one-off lottery ticket – since January 2026, eleven Erdős problems have moved from open to solved, credited to AI, and that’s before you go back to late 2023, when DeepMind’s FunSearch found new constructions for the cap set problem, an open challenge that had kept mathematicians up at night for decades – a genuinely new result that wasn’t in the training data and wasn’t even known. So when you tell me judgment and creativity are the human moat, I have to ask: which part of solving a sixty-year-old conjecture in eighty minutes wasn’t judgment?

Which is the whole point of the fallacy fallacy. The comfortable bet is that this is the loom again, the ATM again, that the displaced work just flows into jobs we can’t picture yet. Maybe. I would love that to be true. But I think it maps to people the same way it maps to skills: 90% don’t get a thrilling new role, they lose the one they had, and 10% become worth more than ever – until their next split. New work, sure. Plenty of it. Just not necessarily for us.

So next time the guy in the back of the room tells you not to worry, ask him one thing. Has he checked with the horses?

Tell me where I’ve got this wrong. I’d genuinely like to be talked out of it.

I hope I am wrong on this one, but the trajectory we are on doesn’t leave me much room for optimism.