Developers increasingly refuse to code without AI tools

This story crystallizes a cultural turning point: for a growing share of developers, working without AI assistants now feels abnormal. METR’s finding that many coders simply refused to participate in tasks if AI tools were withheld is striking because it suggests a shift from AI as an accelerator to AI as a crutch. In the short run, this dependence can look like productivity—it’s easier than ever to scaffold features, refactor code, and generate boilerplate—but it obscures a deeper erosion of debugging discipline, systems thinking, and institutional memory.

For the race to AGI, that dependence cuts both ways. On one hand, widespread adoption of AI coding tools feeds constant interaction data back into the models, accelerating their refinement and making large‑scale “AI writes AI” workflows more plausible. On the other, if human engineers’ baseline skills atrophy, organizations may become less capable of auditing, steering, or even understanding increasingly agentic systems. That asymmetry—stronger models, weaker overseers—is precisely what long‑term safety researchers worry about. The article underscores that governance of AI in engineering teams isn’t just about policy; it’s about maintaining human expertise and incentives in an environment where the path of least resistance is to let the model drive.