AI Made Experienced Developers 19% Slower. They Still Think It Helped.

A randomized controlled trial reveals something unsettling about our relationship with the tools we trust.

The study was designed to be definitive. Researchers at METR recruited 16 seasoned open-source developers, the kind of professionals who have spent years contributing to repositories with over 22,000 GitHub stars and more than a million lines of code. The researchers randomly assigned 246 real tasks from these developers' own projects: bugs, refactors, feature additions. Half the tasks were completed with full AI assistance (Cursor Pro with Claude 3.5/3.7). Half were completed without.

Before each task, developers predicted AI would cut their completion time by 24%. After finishing the study, they reported feeling that AI had reduced their time by 20%. The measured reality: allowing AI increased completion time by 19%.

That is not a small rounding error. That is a complete inversion of expectations, and it is among the most important data points to emerge from the AI revolution so far.

Why Everyone Gets This Wrong

The story we tell ourselves about AI and productivity is seductive in its simplicity. AI assists, humans accelerate. You type a prompt, code appears, you move faster. Millions of developers report feeling more productive with AI tools. GitHub's own data claims Copilot users complete tasks 55% faster in isolated benchmarks. The venture capital flooding into AI coding tools exceeds $10 billion. The narrative is nearly universal: AI makes knowledge workers faster.

But there is a critical flaw in how we measure productivity gains from AI. Most evidence comes from two sources: self-reported impressions (which are notoriously unreliable) and controlled experiments on artificial tasks given to unfamiliar codebases. Neither maps to what an experienced professional actually does in a day.

Real software development on mature, complex systems is not about typing speed. It is about understanding 10 years of architectural decisions embedded in a million lines of code, navigating tradeoffs that live only in the mental model of someone who has spent years in that codebase, and debugging failures that require holding contradictory hypotheses simultaneously. That is not what AI is currently good at. And optimizing for the wrong thing, it turns out, can actively interfere with the thing you were already doing well.

What the Data Actually Shows

The METR study is not alone. A February 2026 working paper from the National Bureau of Economic Research surveyed 6,000 CEOs, CFOs, and senior executives across firms in the United States, United Kingdom, Germany, and Australia. The results were striking in their uniformity: roughly 90% of firms reported that AI had zero measurable impact on either employment or productivity over the prior three years. The executives using AI were averaging just 1.5 hours per week with the tools, and 25% had not used AI at work at all.

Goldman Sachs published its own analysis and found "no meaningful relationship between productivity and AI adoption at the economy-wide level." ManpowerGroup's 2026 Global Talent Barometer found that while regular AI use increased 13% among workers surveyed across 19 countries, confidence in the technology's utility fell 18%. Workers are using AI more, trusting it less, and their employers are seeing no aggregate benefit.

This is not a fringe view. It is now the consensus picture from the broadest, most rigorous data available.

The Deeper Mechanism

Understanding why this is happening requires going back to 1987, when Nobel laureate Robert Solow made one of the most quoted observations in modern economics: "You can see the computer age everywhere but in the productivity statistics."

For 15 years, American businesses poured investment into computers. Productivity growth slowed. Economists argued furiously about whether technology was overhyped or whether the data was simply wrong. Then, in the 1990s, productivity finally surged. The computers did eventually pay off, but only after organizations fundamentally reorganized themselves around the new tools, after workers developed genuine fluency rather than grudging adoption, and after entire workflows were rebuilt rather than just inserting a computer into an existing paper-based process.

The same dynamic appears to be operating now. AI tools are being layered onto existing workflows rather than triggering genuine restructuring. The METR study offers a precise diagnosis: context switching. Experienced developers must shift between deep coding mode and prompting mode dozens of times per hour. Each transition carries cognitive overhead. Traditional IDE work creates flow states; AI-assisted work fragments them. The tool adds friction precisely where skilled practitioners had already eliminated it.

For a junior developer working on isolated, well-defined tasks, AI is a genuine accelerant. The tool fills the gaps. For an expert navigating ambiguity in a complex system, the same tool often introduces gaps where none previously existed.

What This Actually Means

The counterintuitive finding from the METR study is not that AI is useless. Goldman Sachs found 30% productivity boosts for specific, well-defined applications. McKinsey estimates AI could add $2.6 to $4.4 trillion annually to the global economy once adoption matures. Penn Wharton's Budget Model projects a 1.5% GDP boost by 2035. These forecasts are not irrational.

What the data suggests is that we are in the earliest, most disorienting phase of a genuine technological transition. The productivity gains are real in controlled settings. The macro gains have not yet materialized. The gap between those two facts is not a reason for skepticism about AI's eventual impact. It is a description of how every major general-purpose technology has diffused through an economy: slowly, unevenly, and only after the surrounding organizational infrastructure catches up.

The more lasting lesson may be about the perception gap itself. The METR developers did not just perform worse with AI. They consistently believed they had performed better, both before and after the evidence showed otherwise. That is not a failure of the technology. It is a failure of introspection, and it is worth sitting with.

We have spent three years building an enormous cultural intuition that AI equals faster, better, more. That intuition may be self-reinforcing in ways that are genuinely difficult to audit from the inside. A tool that makes you feel more productive while making you measurably less so is a strange kind of trap.

The open question worth carrying forward: if your instincts about AI are systematically wrong in the same direction as everyone else's, how would you ever know?

Subscribe to rkto.substack.com for more.