The AI Takeover Has Completely Backfired
A breakdown of three industries — code cleanup, expert training data, and AI supervision — that emerged precisely because AI failed to fully replace the humans it displaced.
1 / 5at 8:10
The machine that was supposed to make expert knowledge useless just became one of the biggest buyers of expert knowledge on Earth.”
AI Summary
Pooja Dutt argues that the predicted AI takeover of the workforce has backfired, and in doing so has accidentally created three new human-led industries. She examines what went wrong when companies like Klarna and Ford replaced workers with AI, and traces how the failures downstream generated fresh demand for skilled human labour — often better-paid than the jobs that were cut. The three industries she identifies are the cleanup economy, the teaching economy, and the supervision economy.
Yank Summary
What matters, in Yank’s words, each linked to where the source says it.
GitClear analyzed 623 million lines of code changes and found duplicated code blocks up 81% since 2023, while refactoring collapsed from 21% to under 4% of changed lines.
SlopFix charges $10,000 a month for one week of refactoring AI-generated codebases, paid in proportion to how much code they delete.
The machine that was supposed to make expert knowledge useless just became one of the biggest buyers of expert knowledge on Earth.
The central paradox of the source crystallizes here: AI's market hunger for human expertise inverts its original promise of making professionals obsolete.
In context
model-generated data degrades over time. They create more errors. It's basically photocopying a photocopy. Human expert data is the fresh original that keeps information going through the AI models actually accurate, which means that human knowledge isn't just an input in this industry. It's an input that cannot be substituted. So, think about what this means. The machine that was supposed to make expert knowledge useless just became one of the biggest buyers of expert knowledge on Earth. Your radiologist's judgment, your lawyer's opinion, that's now the scarcest commodity of tech. The AI
Receipt
A model trained only on the average internet can only ever be average. To get smarter than the internet, you need data that's better than the internet — and that data doesn't exist online; it exists in the heads of professionals.
Explains why AI improvement has hit a ceiling that only professionals can raise, grounding the source's broader argument about irreplaceable human knowledge.
In context
Overflow. And the models have already absorbed all of that. So a model that's only trained on the average internet can only really ever be average. To get smarter than the internet, you need data that's better than the internet. And that data doesn't exist online, it exists in the heads of professionals. I mean, there's no website where a senior oncologist walks through exactly how she weighs a contradictory scan against a patient's history. There's no forum where a tax attorney shows every step of reasoning through an ambiguous
A model that's only trained only on the average internet can only really ever be average. To get smarter than the internet, you need data that's better than the internet — and that data doesn't exist online; it exists in the heads of professionals.
Receipt
Roles paying $55,000 before layoffs now command $75,000 because they require hybrid humans who can manage, audit, and prompt the AI tools that were supposed to replace them.
Concrete salary data shows that surviving automation has made workers more expensive, undermining the cost-savings case for replacing humans with AI.
In context
actually reaching break even for the first time in May 2026 after all the humans came back. But, the detail everyone's missing isn't that they came back. It's the price of coming back. Roles that were paying $55,000 before layoffs now command $75,000 because they now require hybrid humans who can manage, audit, and prompt the AI tools that were supposed to replace them. So, it's the same job plus AI supervision, which means about a 35% raise. And the cost math underneath that is brutal for the AI is cheaper thesis. So, if you look at the data, cloud compute and API fees scaled into a
Roles that were paying $55,000 before layoffs now command $75,000 because they now require hybrid humans who can manage, audit, and prompt the AI tools that were supposed to replace them.
Receipt
Models only trained on model-generated data degrades over time. It's basically photocopying a photocopy.
The photocopy metaphor establishes why synthetic self-training is a dead end, forcing the source's pivot to why human expert data is structurally necessary.
In context
something that I talk about in the data drip video that I mentioned. So, the obvious cheap solution is synthetic data. Basically, you let the model teach itself and generate the input for its output. But, models only trained on model-generated data degrades over time. They create more errors. It's basically photocopying a photocopy. Human expert data is the fresh original that keeps information going through the AI models actually accurate, which means that human knowledge isn't just an input in this industry. It's an input that cannot
Models only trained on model-generated data degrades over time. They create more errors. It's basically photocopying a photocopy.
Receipt
IBM's HR automation resolved 94% of requests, but the 6% it couldn't handle included ethical judgment calls that a human should have made — and IBM responded by tripling entry-level hiring.
IBM's own automation failure and subsequent hiring surge illustrate that edge cases requiring ethical judgment turn human labor from optional to mandatory.
In context
company goes back and rehires people for more money. So, that last 40% the edge cases and red customers is actually what broke IBM's HR automation. It resolved 94% of requests, but the 6% that it couldn't resolve included ethical judgment calls that a human should have made. And then IBM responded by saying that they're going to triple entry-level hiring in 2026. So, automation was supposed to make human labor optional. Instead, it actually made a specific kind of human labor mandatory. The person who stands next to the machine
IBM's HR automation It resolved 94% of requests, but the 6% that it couldn't resolve handle included ethical judgment calls that a human should have made — and then IBM responded by saying that they're going to triple tripling entry-level hiring.
Receipt
Keep going with this source
Create something
Turn this source into a post, a thread or a newsletter, every line checked against it.
See what you can create →Ask a question
Ask about anything in this video. We’ll try to provide an answer.