Developers Who Refuse to Code Without AI May Be Setting Themselves Up for Trouble

Developers have grown so attached to AI coding tools that many now flatly refuse to write code without them, and researchers are raising serious flags about where that dependency is headed.

The warning signs are piling up fast. AI research lab METR set out earlier this year to repeat a landmark 2025 study measuring how AI affects developer productivity, only to hit an unexpected wall. Developers declined to participate because they were simply unwilling to work without AI, even temporarily, for the sake of a controlled experiment. METR ended up publishing a self-reported survey instead, in which developers estimated that AI made them roughly twice as productive. But self-perception and measurable output are two very different things.

The original 2025 study had already found something uncomfortable: while developers felt more productive with AI coding tools, the tools actually slowed them down overall. Code was generated faster, but engineers spent extra time hunting for errors, correcting AI output, and waiting on completions. The net result was a loss, not a gain.

That pattern is playing out at scale across the industry. Amazon quietly shut down an internal leaderboard called Kirorank that tracked employee AI token usage after workers began gaming the system, running up costs without meaningfully improving output. Uber burned through its entire 2026 AI budget in just four months, and its COO Andrew Macdonald acknowledged on a podcast that the spending had not produced a measurable increase in projects or productivity.

The longer-term concern is code maintenance. Programmer and author James Shore made the case sharply in a widely shared blog post: generating code faster only helps if maintenance costs fall too. If they don’t, developers end up in a worse position, not a better one. Code Rabbit, which reviews open source pull requests, found that AI-generated code produced 1.7 times more problems than code written by humans. Researchers from Singapore Management University published findings in April warning that AI-generated code can introduce long-term maintenance burdens into real software projects.

The trend known as tokenmaxxing, treating token usage as a stand-in for productivity, has dominated developer culture in 2026. But the backlash is arriving quickly. Companies are discovering that heavy use of AI coding tools does not automatically translate into better software or faster delivery.

Even Cognition CEO Scott Wu, whose company makes the AI coding agent Devin, admits that the tool currently performs somewhere between a junior and mid-level developer depending on the task. It is not a hands-off solution. Researchers suggest developers need to understand where AI coding tools succeed and where they fall short, invest in quality assurance systems built around AI output, and continue treating AI-generated code with the same scrutiny they would apply to work from a junior hire. High-stakes tasks like software architecture and security design, researchers and practitioners agree, should remain firmly in human hands.

The dependency on AI coding tools is now so deep that it has become difficult to even study. That alone should give the industry pause.