The research nonprofit METR is overhauling its developer productivity experiments after encountering a significant hurdle: developers simply do not want to code without AI anymore. While a 2025 study famously suggested that AI tools actually slowed down experienced developers by nearly 20%, the landscape has shifted dramatically in early 2026. The widespread adoption of advanced assistants has created a "selection effect," where the most efficient AI users opt out of studies that require them to work "the old-fashioned way," skewing the research results.

This shift makes it incredibly difficult to find a stable control group for comparison. METR researchers noted that participants often avoided submitting tasks where they expected the highest AI-driven gains, fearing the frustration of manual labor if assigned to the non-AI group. Furthermore, the rise of autonomous agents—tools that can execute multi-step coding tasks with minimal human intervention—has complicated time-tracking, as developers often juggle multiple projects while waiting for their AI agents to finish work in the background.

Despite these data challenges, preliminary figures suggest a reversal from last year's trends. For the subset of developers who remained in the study, researchers estimated an 18% speedup. However, METR cautions that the true productivity leap is likely much higher among those who have fully integrated AI into their professional workflows. To fix these blind spots, the organization plans to pivot toward more intensive, higher-paid experiments and fixed-task evaluations to better quantify the accelerating pace of AI-driven software development.

The research nonprofit METR is overhauling its developer productivity experiments after encountering a significant hurdle: developers simply do not want to code without AI anymore. While a 2025 study famously suggested that AI tools actually slowed down experienced developers by nearly 20%, the landscape has shifted dramatically in early 2026. The widespread adoption of advanced assistants has created a "selection effect," where the most efficient AI users opt out of studies that require them to work "the old-fashioned way," skewing the research results.

This shift makes it incredibly difficult to find a stable control group for comparison. METR researchers noted that participants often avoided submitting tasks where they expected the highest AI-driven gains, fearing the frustration of manual labor if assigned to the non-AI group. Furthermore, the rise of autonomous agents—tools that can execute multi-step coding tasks with minimal human intervention—has complicated time-tracking, as developers often juggle multiple projects while waiting for their AI agents to finish work in the background.

Despite these data challenges, preliminary figures suggest a reversal from last year's trends. For the subset of developers who remained in the study, researchers estimated an 18% speedup. However, METR cautions that the true productivity leap is likely much higher among those who have fully integrated AI into their professional workflows. To fix these blind spots, the organization plans to pivot toward more intensive, higher-paid experiments and fixed-task evaluations to better quantify the accelerating pace of AI-driven software development.