METR Study Shows AI Coding Assistants Slow Senior Developers by 19%

AI coding assistants reached mass adoption in 2025, but a controlled study from Model Evaluation & Threat Research (METR) identified measurable productivity losses among senior developers working on mature codebases. The randomized trial analyzed 246 real-world software engineering tasks completed by 16 experienced developers who had contributed to large open-source repositories for multiple years. When AI tools were enabled, developers took 19% longer to complete issues—a significant slowdown that contradicts both developer expectations and expert forecasts.

The research appears in a July 2025 paper published on arXiv, with methodology details available through the METR blog post. Participants primarily used Cursor Pro with Claude 3.5/3.7 Sonnet models—frontier tools at the time of the study. Each task averaged two hours of implementation time, and developers were compensated $150 per hour for their participation.

The perception gap between expected and actual outcomes is striking. Before starting tasks, developers forecast that AI assistance would reduce completion time by 24%. After experiencing the slowdown firsthand, they still believed AI had improved their performance by roughly 20%. The measured reality showed the opposite: tasks took 19% longer with AI assistance enabled. This disconnect suggests developers may not accurately assess their own productivity when using generative tools.

Several measurable factors contributed to the slowdown. AI-generated suggestions required manual verification before integration into production workflows. Developers spent additional time writing and refining prompts to get usable output. Model responses frequently lacked repository-specific context that senior engineers already possessed through years of familiarity with the codebase. Generated code introduced architectural inconsistencies that violated established patterns. Engineers spent time reviewing low-quality outputs, and waiting for AI responses consumed measurable workflow time. Researchers reported that only 44% of AI-generated suggestions were ultimately accepted into production workflows.

Senior engineers working inside mature systems already understand existing abstractions, internal APIs, team conventions, historical design decisions, dependency limitations, and performance bottlenecks. AI assistants lacked direct understanding of those factors in many production repositories. The generated code often matched syntax requirements while violating architectural patterns already understood by experienced maintainers. This is like trying to parallel park a freight train—technically possible, but the friction costs outweigh the benefits.

A separate 2025 academic paper titled "AI-assisted Programming May Decrease the Productivity of Experienced Developers by Increasing Maintenance Burden" identified downstream maintenance costs associated with AI-generated code. Researchers found that less-experienced developers increased output after adopting GitHub Copilot, but experienced maintainers absorbed additional review work. The study measured several effects: senior developers reviewed 6.5% more code after Copilot adoption, original code productivity among core developers declined by 19%, AI-assisted code required additional rework, and maintenance complexity increased over time.

Controlled studies consistently showed stronger AI performance during boilerplate generation, small utility functions, documentation drafting, unit-test generation, and greenfield prototypes. Performance deteriorated during debugging, large-scale refactoring, distributed systems work, multi-service integrations, legacy modernization, and security-sensitive implementation. Research published in 2026 reported that AI assistants reduced completion times by 26–55% primarily in narrowly scoped coding exercises rather than architecture-heavy production engineering.

This distinction matters because senior engineers disproportionately handle production incidents, system design, code review, infrastructure scaling, dependency migrations, and security auditing. AI-generated output frequently required extensive correction in those domains. Google and other large technology companies reported increased AI-generated code adoption during 2025. Internal measurements showed that AI-assisted code frequently required additional review cycles before acceptance. AI-generated pull requests needed more refinement iterations, human reviewers spent additional time validating generated logic, code review duration increased in some repositories, and generated code passed syntax validation more often than architecture validation.

Google reported that AI-assisted code achieved similar acceptance rates to human-written code only after additional review refinement. This effect disproportionately impacted senior engineers because they usually own review authority for production systems. The apparent productivity gains often shifted work from junior contributors to senior maintainers responsible for production quality control (a problem that has plagued engineering teams for years, frankly).

AI tools also generated additional security and reliability risks in production environments. The physical reality of using these tools involves clicking through multiple suggestion iterations, waiting for model responses to load, and mentally tracking which generated snippets actually work versus which need rewriting. The cognitive load of managing AI output often exceeds the cognitive load of writing the code from scratch for developers who already know the system.

Whether organizations will actually reduce AI tool adoption based on these findings remains uncertain. The technology continues to improve, and the study itself notes that progress is difficult to predict. However, the data suggests that for experienced developers working on mature systems, the current generation of AI assistants may create more friction than they resolve. Whether users actually pay for it remains the real question.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn