Study Finds AI Tools Made Experienced Developers 19% Slower — A Cautionary Tale for Rapid Automation

In late 2025, METR (Model Evaluation & Threat Research), a nonprofit that studies AI risks and impacts, published surprising productivity results: experienced software developers were slower when using state-of-the-art AI tools.

METR recruited 16 experienced developers and asked them to complete a total of 246 real-world coding tasks using advanced AI assistance. The study deliberately avoided scenarios like "vibe coding," where less-experienced programmers rely heavily on large language models to produce code for them. Instead, this test measured how skilled professionals integrated AI into ordinary workflows.

Key Finding

Developers using AI tools took 19% longer on average to complete tasks than when working without AI. According to METR and subsequent reporting in Fortune, both researchers and participants were surprised by the result.

Why This Matters

The finding complicates common assumptions that AI will immediately and uniformly boost productivity across white-collar jobs. For software development — a field often considered primed to benefit from automation — the study suggests that integrating AI into expert workflows can introduce overheads: additional verification, context-switching, prompt engineering, or time spent correcting or adapting AI-generated suggestions.

"Many experts have a lot of experience. … We should not just ignore that," said Anders Humlum, an economist who studies AI and workplaces. "I would just take this as a good reminder to be very cautious about when to use these tools."

Broader Context

The METR results arrived amid mounting economic and environmental debates about AI. In 2025, some observers warned of an "AI bubble"—investment in AI far outpacing demonstrable, short-term returns. Investor Michael Burry publicly cautioned that frenzied spending could be inflating expectations.

At the same time, protests and regulatory scrutiny rose over large, resource-intensive AI data centers. Critics pointed to increased local demand for electricity and water, and the U.S. Department of Energy warned about grid strain and a growing risk of blackouts as data-center construction surged.

What Researchers Recommend

METR said it plans to continue measuring AI's real-world productivity effects. In the short term, experts suggest a cautious, evidence-driven approach: pilot AI tools in controlled settings, measure actual time-savings (or costs), and account for infrastructure and environmental impacts when scaling AI deployments.

For individuals, staying informed about data center impacts and contacting policymakers about local infrastructure are practical steps to address broader societal effects.

Bottom line: AI can be a powerful assistant — but for experienced professionals, it is not an automatic productivity multiplier. Careful evaluation and thoughtful integration are essential.