← Back to Posts

What Marx Gets Right About AI at Work

I've been reading Marx lately (I know, I know). Not because I've turned into a revolutionary. I've spent my whole career in startups and game development, and these days I'm doing AI-assisted marketing engineering. I like markets, I like shipping, and I like owning what I build.

Over the past year of code work, I've noticed that ideas that used to be huge time sinks (and maybe not worth the effort) are now almost free. A focused weekend can replace three months of design and dev work on something complex, as long as it's well-known enterprise territory. It's often cheaper to throw the old thing away and "code" it from scratch than to fix what already exists. Crazy times.

But Marx asked a question that most of the AI commentary I read just skips: when a machine does the work, who gets the gains? I think that's THE question for knowledge work right now. So here's a handful of thoughts on what he got right, what he got wrong, and a few ways this could go well. Hopefully it's useful to anyone else building in this space.

Marx TFL;DFR

Capital was a slog. I hated reading it. It could have been a solid pamphlet. But it's mostly an analysis of how capitalism works (there's almost no blueprint in there for what comes after). Three ideas matter here:

  1. Surplus value. Workers produce more value than they get paid in wages. That gap is where profit comes from.
  2. Machines replace labor. Competition pushes every firm to automate, whether the owner wants to or not. Fall behind on cost and you're done.
  3. Ownership decides who wins. Whoever owns the tools captures the productivity gains. If all you own is your time, you don't have much leverage.

He also spotted a trap. Every firm cuts labor to stay competitive, but workers are also customers. Cut enough wages across the whole economy and there's nobody left to buy your stuff.

What lines up with 2026

  1. The automation arms race. Nobody I know adopts AI because they want fewer coworkers. They adopt it because the other guys already did. Marx called this coercive competition, and I think it's a pretty fair description of most roadmap conversations happening right now.
  2. The bottom rung breaks first. Stanford researchers dug into ADP payroll data through June 2026. Employment for 22-25 year olds in the most AI-exposed jobs fell about 11% since late 2022, while the least-exposed jobs grew about 10% (Brynjolfsson, Chandar, and Chen). The drop comes from fewer hires rather than more layoffs. The authors are careful to call these early signals and not proof of cause (fair enough).
  3. The gains could flow to owners. Labor takes home about 60% of US income today. Anthropic's own economic scenarios show that dropping to 45.2% in the most extreme case, with knowledge-worker wages down 11.5% against a no-AI path (VKTR coverage). They don't put odds on any scenario. Still, it's kind of wild when an AI lab publishes a chart Marx could've drawn.
  4. Our shared knowledge became somebody's asset. In an 1858 notebook, Marx predicted that society's accumulated know-how (he called it the "general intellect") would become the main engine of production. LLMs are basically that idea made literal. They were trained on what millions of us wrote, and a handful of companies own them. US private AI investment hit $285.9 billion in 2025 (Timcke, citing Stanford's AI Index), so the buy-in is NOT small.

What he got wrong

  1. The value theory doesn't survive automation. Marx held that only human labor creates new value. Follow that logic and a nearly labor-free company should earn nearly nothing, which is a problem even Jacobin's take has to work around. Anyone who's run a software business knows that code, capital, and ideas create value too.
  2. The collapse hasn't shown up. That same Stanford paper finds no economy-wide job displacement. Overall employment in their sample rose about 6%, and workers over 40 in exposed jobs are doing fine. Goldman Sachs reports no significant AI-driven shift in the overall US job mix so far.
  3. He underrated new work. Every past automation wave created jobs nobody could've named in advance. My current job title didn't exist five years ago. Marx saw the destruction clearly and mostly missed the creation.
  4. Wrong class. Marx was writing about factory hands. This wave lands first on people with graduate degrees and equity grants, and plenty of them are part owners of the companies doing the automating.

Four scenarios I'm optimistic about

Marx's core complaint was that workers got separated from the tools of production. I think each of these is a different way of closing that gap.

  1. The indie studio economy (and my favorite). Game development already ran this experiment. Shipping a game used to take a publisher, a big team, and a retail deal. Then engines like Unity and Unreal got cheap, digital storefronts opened up, and small teams (sometimes one person!) started shipping hits. AI can do the same thing for knowledge work. A marketer/analyst/designer with good models can do what used to take a department. In Marx's terms, the worker owns the means of production again. This only works if model access stays cheap and competitive. If three companies control the only engines that matter, it's sharecropping.
  2. The gains come back as time. Marx's most hopeful passage argues that real wealth is free time, not output. Keynes made a similar bet in 1930 when he predicted a 15-hour work week. If AI makes a team twice as productive, one option is half the team. Another is the same team on a four-day week. Which one we get is a choice (somebody just has to go first without getting undercut).
  3. More people own a piece. Startups already solved a version of Marx's problem. We give employees equity so the people doing the work share the upside. If the gains flow to capital, make more people capitalists. That could be broader employee ownership, public stakes in AI infrastructure, or dividends for the people whose data trained the models. It'll probably take more than goodwill though. The authors of a recent paper called "The AI Layoff Trap" argue that worker equity alone doesn't stop the race to over-automate, and they favor a tax on automation instead (abstract quoted by John Ganz).
  4. The ladder gets rebuilt (and it's shorter). The junior squeeze is the most worrying data point. I also think it's the most fixable. The Stanford study found the losses where AI automates tasks, and flat or rising employment where it complements people. An AI tutor that never gets tired could maybe compress five years of apprenticeship into two, so juniors get to the hard problems sooner. The World Economic Forum's employer survey still projects 170 million new roles against 92 million displaced by 2030 (summary). But companies have to keep hiring juniors while the old entry-level tasks disappear, and that's a design decision every team lead makes.

What makers can do

None of the good outcomes are automatic (and none of the bad ones are either). The people designing these systems are making the call right now, one workflow at a time.

  1. Build to complement. The data says the damage sits where AI replaces a person and the growth sits where it extends one. Try to design for the second.
  2. Keep a bottom rung. If AI does the junior work now, give juniors the review, the judgment calls, and the client contact earlier.
  3. Spread the ownership. If a tool lets five people do the work of fifty, those five should own more of the result.
  4. Watch your own demand curve. Your customers are somebody's employees.

You don't have to buy Marx's conclusions to steal his question about who gets the gains. Those of us building this stuff get to answer it, and I'd like the answer to be "more of us than last time."

Sources