They Let AI Write Their Code, but Not Judge Their Hires

Google DeepMind's AGI Safety and Alignment team — the people whose job is to make sure AI can be trusted — has built a special door around its own company's hiring process. Applicants are handed a form that routes around Google's AI screening, with a warning about what the system might do to their applications. ^[1]

This team set up a special form to go past the recruiter review, and get their resumes direct to the people on the team. — DeepMind

The form also asks applicants to write their responses themselves, rather than let an AI do it.

We have an applications system with a non-trivial probability your CV will be screened out incorrectly or take too long to reach us. — Google DeepMind AGI Safety and Alignment Team

In the past month, the loop has stopped being a projection and become a description. These are the same labs feeding their own work back into themselves. Claude now writes 80% of Anthropic's code, with an eightfold increase in code produced per person, and OpenAI's GPT-5.3 Codex is contributing to its own development end to end. ^[2] DeepMind has built a system that uses AI to clean the data that trains AI. ^[3] The machines are increasingly making the machines. And the labs' own behavior says they know the loop can't be trusted to judge. A randomized trial from METR found developers working with AI were 19% slower than those without it, despite predicting they'd be faster. ^[4] A survey found 96% of developers don't fully trust AI-generated code. ^[5] The maintainers of the Godot game engine, drowning in AI-generated pull requests, say filtering the AI output often takes longer than writing the fixes themselves. ^[6]

If you want to help, more funding so we can pay more maintainers to deal with the slop (on top of everything we do already) is the only viable solution I can think of — Rémi Verschelde

The CEOs are not hiding what comes next. Anthropic's Dario Amodei says he can see a time when the company will need fewer people at the junior and intermediate levels. ^[7]

I can see it within Anthropic, where I can look forward to a time where on the more junior end and then on the more intermediate end we actually need less and not more people. — Dario Amodei

DeepMind's Demis Hassabis confirms a slowdown in hiring at the junior level. ^[7]

Now I think maybe we're starting to see just the little beginnings of it, in software and coding. — Dario Amodei

It is already happening. Stanford's study of the entry-level market found the employment gap for young workers widened from 15% to 19% in a single year, driven by reduced hiring rather than layoffs. ^[8] The money is still moving, but only at the top — OpenAI and Meta are spending billions to poach each other's senior engineers, not to grow new ones. The contradiction closes on itself. The labs are automating away the entry-level engineering work — the code-writing, the bug-fixing, the grunt review — where human judgment was supposed to be built. That judgment is exactly what the verification evidence says the automation still needs: someone has to catch the plausible-but-wrong code, the test that passes only because the same flawed assumptions wrote both sides. The people who would have learned to do that catching are the junior engineers the labs no longer need. What breaks is the apprenticeship pipeline — the years of small, supervised tasks that turned new graduates into the senior reviewers the labs still have to hire from each other at any price. The DeepMind form is the confession that the pipeline is already gone: the lab building AI to replace human judgment went around its own AI to find the humans it still trusts to judge.