Here's an inconvenient truth about AI coding assistants: they might actually be making experienced programmers slower. That's the surprising conclusion from a 2025 experiment that should make anyone hyping AI productivity gains think twice.

When AI "Help" Becomes a Hindrance

Nonprofit research group Model Evaluation and Threat Research (METR) ran an experiment in July 2025 with 16 software developers who averaged five years of experience. These weren't beginners—they were seasoned professionals. The developers tackled 246 tasks total, splitting them evenly between AI-assisted work and traditional coding.

The AI tools in question included Cursor Pro and Claude 3.5/3.7 Sonnet, which represent the cutting edge of coding assistance technology.

Going into the study, developers expected AI to shave 24% off their task completion times. Instead, the opposite happened: AI-assisted tasks took 19% longer than doing the work the old-fashioned way.

Participant Philipp Burckhardt captured the disconnect perfectly in a blog post: "While I like to believe that my productivity didn't suffer while using AI for my tasks, it's not unlikely that it might not have helped me as much as I anticipated or maybe even hampered my efforts."

The Hidden Tax of AI Code

So what went wrong? According to researchers, the problem boils down to context—or rather, AI's lack of it. Study author Nate Rush explained that while AI generates outputs that seem generally useful, "developers have to spend a lot of time cleaning up the resulting code to make it actually fit for the project."

Developers found themselves spending extra time debugging AI outputs, adapting generated code to their specific project needs, crafting effective prompts, and simply waiting for AI responses. The promised efficiency gains got eaten up by quality control.

The Broader AI Productivity Question

This finding fits into a larger pattern of mixed results on AI's workplace impact. Anthropic's own study of Claude Code found that while workers completed more tasks and delegated up to 20% of repetitive work, there were downsides: collaboration declined, skills eroded, and employees worried about becoming obsolete.

Meanwhile, a study of 25,000 Danish employees using AI chatbots between 2023 and 2024 found essentially no meaningful changes in wages or hours worked. The productivity revolution keeps getting promised but hasn't quite materialized.

The stakes matter. Sen. Bernie Sanders (I-VT) has warned that AI should benefit workers broadly, not just billionaires, pointing to predictions that half of entry-level white-collar jobs could vanish within five years. If AI isn't even delivering productivity gains yet, those displacement risks look even more troubling.