Authoring Is Cheap. Verification Is the Constraint.

This is the single most important finding in the field, and the one most often misquoted. Faros AI published it first in July 2025 from telemetry across more than 10,000 developers on 1,255 teams. Tasks completed per developer were up 21%. Pull requests merged per developer were up 98%. PR review time was up 91%. Average PR size was up 154%. Bugs per developer were up 9%. And at the company level - throughput, DORA metrics, quality KPIs - there was no significant correlation with AI adoption at all. Twice the PRs, and nothing moved.

Faros ran it again in May 2026 across 22,000 developers and 4,000 teams, and the picture got worse, not better. Tasks per developer up 33.7%. PR merge rate up 16.2%. PR size up 51%. Bugs per developer up 54%, from 9% a year earlier. Incidents per PR up 242.7%. Median PR review time five times longer. Code churn up 861%. Deployments per week down 11.7%. PRs merged with no human review at all up 31.3%, with a quarter of all PRs now reviewed by an AI agent instead of a person.

DORA has been tracking the same shape from the survey side. In 2024, a 25% increase in AI adoption was associated with a 1.5% drop in delivery throughput and a 7.2% drop in delivery stability. In 2025, throughput finally turned positive while stability stayed negative, and DORA settled on the word that explains both: AI is an amplifier. It magnifies whatever is already there. In 2026, DORA's ROI report named the two mechanisms - a J-curve where productivity dips before it rises, and a verification tax, the review and validation cost that cheap authoring shifts onto humans.

Then there is METR, which did the thing nobody else did: a randomized controlled trial. In July 2025, 16 experienced open-source maintainers worked 246 tasks on their own repositories. With AI tools they were 19% slower. Before the study they had forecast being 24% faster. After the study, having just been measured at 19% slower, they believed they had been 20% faster. METR's February 2026 update with late-2025 tools found the original cohort still estimated at minus 18% (confidence interval minus 38% to plus 9%) and a fresh cohort at minus 4%, and METR itself called the result "very weak evidence" and is redesigning the study because developers now refuse to work without AI.

Take two things from METR and not a third. First, perceived speed and measured speed diverge, by about 40 points, in the same people, on the same work. Second, the effect on experienced developers in mature codebases is not settled. Do not take from it that AI makes engineers slower. Take from it that your engineers' own report of how much faster they are is not data, and if that is what your dashboard is built on, you do not have a dashboard.

Organizational performance is not the sum of individual output. It is how well code integrates, how reliably review catches defects, and whether the architecture stays coherent while forty people and a fleet of agents push at it. Speed is the most visible signal in that system and the most misleading one. When PRs per developer are up and deployments per week are flat or down, you do not have a gain. You have a queue, and the queue is where the money went.

The dangerous failure mode is not one incident. It is quiet accumulation: an inconsistent quality bar across teams, ungoverned data flows, shadow tooling leadership cannot see, and a widening gap between how productive everyone feels and how healthy the system actually is. None of it shows up on a velocity chart. The bill arrives later, as a production break, a security audit finding, or the morning a routine change turns out to need a rewrite.

GitClear's January 2026 data, from 623 million code changes indexed to 2023, is what that tipping point looks like in aggregate. Code-block duplication up 81%. Error-masking constructs - empty catches, broad excepts - up 47%. Cross-file function calls down 35%, which means generated code stopped reusing what already existed. Refactoring's share of all changes fell from 21% in 2022 to 3.8% in 2026. GitClear's own framing is the right one: the headline is not that AI writes bad code. It is deferred structural maintenance, at scale, with nobody assigned to it.