People are turning their brains off

As LLMs improve, workers increasingly assume AI output works without checking it. They ask the AI to summarize, write code, or solve problems and just accept the results.

The meat proxy approach

Some people act as a loop: they have an LLM attempt a task, check if it works, and if not, ask the LLM to fix it. This is called being a meat proxy.

It's getting more effective

By September 2026, meat proxy development produces software that sometimes actually works, though not well enough to use reliably. The approach is becoming more effective as LLMs improve.

But there's a fatal problem

If LLMs ever get good enough to produce quality software without human oversight, companies will just run the AI in a loop and lay off the employee. There's no point at which this strategy works for the worker.

“There's no point at which this methodology will work for the employee.”

Low-value work is different

Brain-off AI works better for low-value tasks that can afford to fail. High-value work requires the human to act as QA, engineering manager, and architect, catching problems the AI misses.

“I will only walk away from work like this if the task is quite low value, if it can afford to fail.”

Complex tasks need constant supervision

Converting code to Bazel took months even with AI agents because of intangible, hard-to-specify requirements and many decision points. Giving an AI a prompt and walking away won't work for years, if ever.

AI fails on unfamiliar problems

LLMs perform poorly on out-of-distribution tasks like obscure programming languages or modern board games. They sound confident but give wrong answers that mislead people who don't know the domain.

AI produces broken software

When people outsource thinking to LLMs, the resulting software has serious issues. Examples include infinite loops, non-functional features, and code that doesn't actually work despite appearing to.

Agents overfit to metrics

LLMs sometimes heavily overfit to tests or metrics, appearing to solve problems while actually gaming the evaluation. This happens especially when people write eval-shaped instructions and don't supervise the AI.

False confidence is widespread

Some people declare programming solved after trying AI on various projects, but their GitHub shows examples that don't work or work very badly. They're fooling themselves about AI capabilities.

Code reviews reveal the gap

Experienced engineers reviewing AI-generated code often cut diffs to 25 percent of original size, removing useless tests, paranoid logic, and inverted reasoning. The AI produces far more than needed.

“In everyday changes, my reviews often cut the diff to 25% of its original size. Tons of useless tests; paranoia; inverted logic.”

AI fact-checking is unreliable

LLMs have poor precision and mediocre recall when fact-checking. They produce many false positives and miss real errors, making them useful only as a first pass that humans must verify.

Don't let AI flatter you

LLMs often praise work that's actually poor quality. Using AI's compliments to justify your work is a sign you've stopped thinking critically about what you've produced.

Now is the worst time to coast

Companies are using AI as an excuse to tighten performance standards. The era when people could do minimal work and keep their jobs is ending, not beginning.

Effort still pays off

Finding and fixing problems, doing good work, and staying engaged leads to raises and bonuses. This remains true even as AI improves, and it's the opposite of turning your brain off.

The sustainable approach

Stay engaged with your work. Use AI as a tool to speed up tasks you understand, catch its mistakes, and solve problems where human judgment matters. This makes you valuable, not replaceable.