In February 2026, AI research lab METR encountered a significant obstacle while attempting to update earlier research on AI coding productivity. The researchers found that most developers simply refused to participate in their follow-up study because they "did not wish to work without AI," even on a limited number of tasks. This reluctance to code without AI assistance marks a stark shift in developer behavior, according to METR.
The original research from 2025 measured how much time open source developers spent on tasks manually versus with AI assistance. While developers reported feeling more productive with AI, the actual results told a different story. AI tools did generate code faster, but developers then had to spend additional time finding and fixing errors, carefully directing the AI, and waiting for it to complete tasks. The net effect was that AI actually slowed developers down rather than speeding them up.
METR's May 2026 survey attempted to capture productivity gains through self-reported data from technical employees. Unsurprisingly, developers perceived that AI tools made them roughly twice as valuable to their organizations. However, recent corporate incidents cast serious doubt on such self-assessments. Amazon was forced to shut down its internal token-tracking leaderboard called Kirorank after employees began gaming the system by using AI agents excessively, running up enormous costs in the process. Similarly, Uber exhausted its entire 2026 AI budget within just the first four months of the year, and COO Andrew Macdonald acknowledged on a podcast that this spending had not produced any measurable increase in project output. These examples suggest that while developers have become dependent on AI tools, that dependence does not automatically translate into actual productivity gains.