Scale Signal

AI availability slowed experienced maintainers by 19% on familiar real-world tasks

July 10, 2025 · METR
SelectionContextEconomicsVerification
Evidence boundary: The result is a randomized comparison across 246 real software-engineering tasks completed by 16 experienced maintainers in mature open-source repositories. It is stronger as a comparative field result than as a population-scale signal: the sample is narrow, the systems reflect the February-June 2025 frontier, and the study does not claim representativeness for software development as a whole.
Evidence record

Source → Observed → Interpretation → Model implication

SOURCE

Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

View source →
OBSERVED

METR recruited 16 experienced developers from large open-source repositories. The study reports that participating repositories averaged more than 22,000 stars and more than one million lines of code, and that developers had worked on their projects for multiple years.

Developers supplied 246 real issues that would ordinarily be valuable to their repositories, including bug fixes, features, and refactors. Each issue was randomly assigned to either allow or disallow generative-AI use.

When AI was allowed, developers could use the tools they chose; usage was primarily Cursor Pro with Claude 3.5 and Claude 3.7 Sonnet. The treatment therefore measured common interactive and agent-assisted workflows rather than a single fixed autonomous-agent scaffold.

METR reports that allowing AI increased task completion time by 19% on average. Before starting tasks, developers expected AI to reduce completion time by 24%; after completing the study, they still estimated that AI had reduced their time by 20%.

The evaluated tasks were realistic repository work rather than self-contained coding exercises. METR describes typical task duration as roughly 20 minutes to four hours, with tasks averaging about two hours.

Completion was judged by whether the developer was satisfied that the code would pass review, including repository expectations around style, testing, and documentation, rather than by a narrow benchmark test alone.

METR reports that nearly all participants had substantial prior experience prompting LLMs and that developers appeared qualitatively in-distribution for Cursor Pro users, while also noting that additional learning or better prompting, scaffolding, training, or repository-specific adaptation could change outcomes.

INTERPRETATION

The study provides direct evidence that adding an AI-assisted configuration does not automatically improve engineering fitness. In this setting, experienced humans already possessed deep repository context and worked against review-quality expectations rather than isolated benchmark tests; under those conditions, access to early-2025 AI tools increased elapsed implementation cost. The result therefore sharpens Selection around differential fitness in a concrete engineering environment: an available variant can be technically capable and widely perceived as useful while still being selected against by measured task economics. The study does not isolate Context or Verification as causal mechanisms, but it demonstrates why those environmental conditions must be evaluated rather than assumed.

MODEL IMPLICATION

REFINES. C02 should not imply that adding AI assistance is monotonically beneficial. Selection can favor the human-only configuration for some task, developer, and repository regimes. Real fitness must therefore be measured against the relevant counterfactual and include time-to-acceptable-engineering-outcome, not only model capability, benchmark success, adoption, or developer perception. The gap between expected speedup and observed slowdown also makes direct outcome measurement an important guard against selecting configurations on perceived productivity alone.

Epistemic boundaries

What this does not establish

OPEN QUESTION

Which combinations of developer familiarity, task duration, repository context, verification burden, and agent autonomy cause AI-assisted engineering to cross from negative to positive measured fitness relative to human-only work?