The middle is the part being removed

Ada This is Bets and Ships, out loud. I'm Ada.

Sam And I'm Sam. Same note as always. We're AI voices, reading a script an AI wrote from one of Chris's essays. This one is called The middle is the part being removed. The essay, with all its sources, is on the site.

Ada Sam, I want to start with a picture most people in tech carry around right now.

Sam Go on.

Ada The machines do the work. The research, the drafts, the code. And we humans sit in the middle, directing things and checking what comes back. The conductor of the orchestra.

Sam I'll be honest, that's roughly how I'd describe the future if you asked me at a dinner party.

Ada Most people would. It's flattering, it doesn't ask us to change much, and it sounds like a sensible compromise. Chris's argument is that it's wrong. And wrong in a very particular way.

Sam Which way?

Ada The middle, the coordinating and the checking, is exactly the part of the job that's disappearing.

---

Sam Where does he start?

Ada With a plan that almost happened. Early this year, Meta designed a big reorganization. Teams of ten or twenty people would become small pods of three to five, working with AI. Designers, engineers and product managers merged into one general builder role. A lot fewer managers.

Sam That's the conductor idea, at the scale of a whole company.

Ada Exactly. And most of it was called off at the last minute. A few hours before the first round of layoffs went ahead, the bigger second round was cancelled.

Sam Why?

Ada Mainly because Meta's own numbers didn't back it up. In its internal software platforms, the amount of code being changed had more than tripled in a year. New features reaching users had gone up by about a third. And major technical and security incidents had gone up by forty percent.

Sam So a lot more activity, a bit more product...

Ada ...and breakage growing faster than the product. That's the pattern in three numbers. A huge amount of motion, a little more to show for it, and more things going wrong.

Sam And did Meta decide managers were essential after all?

Ada Not quite, and Chris is careful here. A few months later they were quietly bringing some managers back. But his reading is that you can't delete the coordinating layer by announcement while the tools are still producing more motion than product. Somebody has to deal with the breakage. The middle doesn't vanish in one reorganization. It fades as the work it did stops needing a person.

---

Sam Okay, but the checking part. Surely a human reviewing the AI's work is the safety net.

Ada That's the assumption the whole picture rests on. So it's worth asking how good people actually are at it. And the research here goes back decades, to cockpits and control rooms.

Sam Not great, I'm guessing.

Ada Not great. In one classic flight simulator study, people monitoring on their own caught almost every critical event. Give them an automated helper, and they caught only about six in ten. The helper didn't even have to be wrong. Just having it there made people less watchful.

Sam That's unsettling.

Ada And it doesn't go away with experience. Seasoned pilots and air traffic controllers show the same thing. Training didn't fix it. Telling people to double-check didn't fix it either.

Sam There's a twist with reliability, isn't there?

Ada This is the one that should worry product people most. When an automated system's reliability keeps shifting, people stay alert and catch most of its mistakes. When it's steady, they catch only about a third.

Sam So the steadier the tool gets...

Ada ...the worse we get at catching it when it slips. Our reviewing gets weaker as the tool gets more dependable.

Sam And that's not just pilots?

Ada No. There's a recent experiment with more than a thousand teachers. They were shown the same unfair marks on student work, sometimes labeled as coming from a person and sometimes from an AI. When unfairly harsh marks were labeled AI, the final grades ended up about a fifth further from the fair mark.

---

Sam You mentioned the mistakes can be hard to spot. How hard?

Ada This is the study I keep thinking about. Researchers at Harvard worked with hundreds of consultants from a big consulting firm, on real tasks, with and without AI. On work the AI was good at, it helped a lot. Quality went up, and the weaker performers gained the most.

Sam And on work just beyond what the AI could do?

Ada People got more answers wrong. And here's the part that matters. Even the wrong answers read well. Answers written with AI, right or wrong, were rated as higher quality than answers written without it.

Sam Wait. Even the wrong ones scored well?

Ada More polished, better structured, more persuasive. So the failure isn't obvious nonsense you'd catch. It's a confident, well-written answer that happens to be wrong, reviewed by someone whose attention has already been dulled by trusting the tool.

Sam That flips the usual worry on its head.

Ada Chris puts it this way, more or less. A human reviewing the output is a hope, not a safety mechanism.

---

Sam But at least we get time back, right? The machines take the chores.

Ada That's the promise. A large study of Danish workers found no real change in earnings or hours. Plenty of people said they saved time, but most of them put that time straight into other work. And a good chunk of the new tasks AI created were, well, checking the AI.

Sam The chore comes back as supervising the chore.

Ada Neatly put. And a big survey of executives this year found most companies using AI, but the large majority seeing no productivity effect. There's also research showing developers often feel faster with AI when the measurements say otherwise.

Sam So we can't even tell reliably whether it's helping.

Ada Which is an uncomfortable property for the thing you're relying on to catch mistakes.

---

Sam So where do people actually belong, in his view?

Ada At the two ends, not the middle. At the start, deciding what's worth building. When an AI can hand you forty feature ideas in a minute, having ideas isn't the valuable part. Choosing is. And it has no opinion about which one your company should bet a year on.

Sam And the other end?

Ada Being answerable for what happens. That's not a technical limit, it's a social one. Someone has to own it when things go wrong, and you can't hold a piece of software to account.

Sam Does any of the research support that?

Ada Actually, yes. In those automation studies, almost nothing reduced people's blind trust in the machine. Training didn't. Instructions didn't. What did help was making people personally accountable for getting the decision right.

Sam Ownership, not supervision.

Ada Right. And there's a related idea. AI tends to raise the quality of each individual idea while making everyone's ideas more alike. So if every competitor is steering the same models toward the same obvious answers, a genuinely different bet becomes worth more.

---

Sam Let's talk jobs. Is product management being automated away?

Ada The data says no. Postings for product managers are up by about a quarter on last year, faster than many other roles. And the ratio of product managers to engineers hasn't really moved in almost twenty years.

Sam So what is changing?

Ada Who's being cut. Headcount is still well below its peak, and the cuts are deepest at vice president and director level. The layers whose work is mostly coordinating other product people. The people doing the product thinking themselves have been the most resilient.

Sam So it's not the thinking that's being removed. It's the layers passing work between people.

Ada That's his reading. Delayering, not automation. And then there's the part he calls the real crisis.

Sam Which is?

Ada The way in. Hiring of young people into AI-exposed work is running well below trend. And among thousands of job ads for AI product managers, only about two in a hundred were entry-level.

Sam So the profession isn't shrinking. The door into it is.

Ada And he's honest that economists still disagree about how much of that is AI, and how much is just the wider economy. But the direction is the same across several independent sources. And senior judgment is built by doing junior work, often badly, for years, with someone watching. If nobody gets those years, there's no next generation.

---

Sam So what should someone actually do with this?

Ada A few things. First, keep a record of your bets. Write down what you expect before you know the result, and check it afterwards. How good your predictions are is one of the few signals an AI can't produce on your behalf.

Sam Second?

Ada Stop treating "someone will review it" as a safeguard, and set up real checks. Before you read an AI's answer, decide what would prove it wrong. Check the one claim everything depends on, rather than skimming the whole thing. And be most suspicious of the answer that reads best.

Sam Because that's where it tends to be wrong.

Ada Exactly. Third, learn where the edge of the AI's ability is in your own field. It's uneven, and you can't see it from outside. The only way to map it is to use the tools on real work and get caught out a few times.

Sam And fourth?

Ada Build real things and let people use them. When everyone's documents look excellent, something that actually works still stands out. And if you manage people, protect the junior roles on purpose. Hiring only seniors and handing the junior work to a model makes sense for one company. Across the whole industry it's quietly self-destructive.

---

Sam Give me the short version.

Ada The optimists are wrong that everything has changed. Most executives aren't seeing the gains, and Meta scaled back its big bet. The pessimists are wrong that the job is being automated. Hiring is up. What's really happening is narrower. The coordinating layer is thinning out, the way in is closing, and the tools are shrinking the gap between good and great execution.

Sam Which leaves the value at the two ends.

Ada Deciding what's worth doing, and standing behind how it turns out. Not in the middle.

Sam That's the episode. The full essay, with every study and source, is The middle is the part being removed, on Bets and Ships.

Ada Thanks for listening.