Coders are refusing to work without AI — and that could come back to bite them

This TechCrunch article focuses on the heavy reliance of software developers on AI coding tools in 2026 and the long-term engineering risks this reliance may bring. The article points out that AI can indeed help developers produce code faster, but researchers warn that faster does not mean better, nor does it mean overall productivity has truly improved. The AI research organization METR previously measured the time open-source developers needed to complete tasks with and without AI assistance, and the results contradicted the developers' own perceptions: they thought AI made them more efficient, but in the actual tests, AI may instead have slowed down the work, for reasons including the time spent correcting AI-generated errors, steering the model, waiting for output, and repeatedly checking results. METR wanted to redo the experiment in 2026 to observe changes after models improved and developers became more proficient, but ran into a new problem: many developers were unwilling to temporarily stop using AI during the research, making the originally intended controlled experiment difficult to conduct. The article also questions the practice of companies measuring AI productivity by token consumption, mentioning that Amazon's internal token leaderboard Kirorank was shut down because employees overused AI agents and drove up costs; Uber was also reported to have used up its AI budget in the first four months of 2026, with executives saying the related spending had not yet brought measurable project or productivity gains. Beyond the cost issue, the article is more concerned with the maintenance burden. Programmer James Shore argues that even if AI doubles the speed of writing code, if maintenance costs do not fall correspondingly, the team is merely trading short-term speed for long-term debt. The article cites multiple sources noting that AI-generated code may bring more bugs and code-review pressure; although some of the data comes from companies selling AI code-review tools and thus carries a commercial bias, research from Singapore Management University also warns that AI-generated code may introduce long-term maintenance costs in real projects. As for solutions, the article does not advocate stopping the use of AI, but emphasizes that developers need to understand which tasks AI is good and bad at, just as they understand a programming language, and build stronger quality-assurance processes. Even AI coding agents like Devin, which can help fix and handle part of the work, still have capabilities closer to that of a junior to mid-level engineer and cannot be treated as a fully hands-off replacement. Architectural design, security design, and critical judgment should still be the responsibility of humans.