David Graeber had a theory. A lot of white-collar work is quietly pointless, and the people doing it suspect as much. He called them bullshit jobs, and he wrote a whole book about it.

He sorted them into types. The ones I keep thinking about: box-tickers, who produce the paperwork that proves work is happening, and taskmasters, who supervise people who don't need supervising. The layer Graeber was suspicious of. Middle management, roughly.

Not every manager I worked with fits that mould, mind you. Some were excellent, some less so. Just like developers, come to think of it.

Fifty years in software teaches you one thing down among the code: the people doing the work are the ones who know the work. The expertise lives at the bottom. The layers above translate and coordinate. Useful, sometimes. Pointless, sometimes.

Reading Graeber, I had an uncomfortable thought. What if these jobs stop being bullshit? Not because anyone redesigns them, but because AI hands the people in them something to actually build with. A product manager who can vibe-code a working app isn't just coordinating any more. They're producing.

That's good news for bullshit jobs. It's bad news for developers. If the coordination layer can build the thing itself, it doesn't need the layer underneath it. The bullshit jobs get less bullshit by swallowing the real ones.

I want to be careful here. I don't see this happening yet, not at any scale that would matter. But I've been retired for over a year, watching the trade from the outside now instead of from inside a dev team. Call this a risk I see coming, not a report from the front.

And there's a complication that cuts the other way. Developers are leaning on AI just as hard. A lot of the ones I know do. Some let it write every line, start to finish. Me too, in my hobby projects. I already wrote about that earlier.

So the real line was never "developer versus manager." It's who has the gut feeling, and who doesn't.

I spent fifty years building software, and that instinct is still with me. Even in hobby code, where nothing's at stake, I catch the wrong turns without trying. When the AI writes something wrong, I can smell it. When it steers us into a dead end, I notice before we're stuck. The stakes aren't what make the shortcut safe. The nose is.

Hand the same tool to a product manager who never learned to program, and there's no instinct to catch the wrong turn. You get software nobody in the building understands. Built by the layer that used to just manage it.

And that's the answer to the question in the title. AI doesn't end the bullshit job. It promotes it. The paperwork that proved work was happening becomes software that nobody asked for and nobody understands. Same pointlessness, better production values.