Your Team Got 66% Faster. Nobody Priced the Bill the System Just Mailed Downstream.
July 1, 2026 · Kevin Kelly, Out of Control~6 min read
Your team just got 66% faster, and that is precisely the problem. Not the tired kind of problem — the kind where the speedometer reads great while the engine quietly throws a rod. In June a study of roughly 22,000 developers put a number on the thing every senior engineer had started to feel in their gut: AI is shipping more code than ever, and the bill for it is coming due somewhere the dashboard doesn't look.
The report that named the bite
Faros AI called it Acceleration Whiplash, and the phrase is doing honest work. Across those ~22,000 developers, AI lifted raw throughput by about 66%. Wonderful. Then the same data showed incidents climbing faster than throughput, not in step with it but ahead of it, and the gap widening precisely where AI adoption ran deepest. Opsera's 2026 report fills in the mechanism: AI cut time-to-PR by as much as 58%, but those AI-authored pull requests then sit in review roughly 4.6× longer and carry an estimated 15–18% more security vulnerabilities. Teams leaning hardest on AI finish about 21% more tasks and merge 98% more PRs, while their review time swells 91%. Faster in, slower through, more to clean up later.
You didn't buy speed — you borrowed it
Here is the sleight of hand worth naming. The wins are all measured at the front of the pipe, where they are loud and legible: lines written, PRs opened, tasks closed. The costs land at the back, where nobody is holding a stopwatch: the review queue, the incident channel, the security backlog, the six-month-out maintenance nobody scheduled. Call it cost displacement. The output got cheaper; the work did not. It just moved downstream, to a place your metrics don't watch. The cheap velocity wasn't bought; it was borrowed. Think of it as a throughput mortgage, and the interest shows up as incidents outpacing the very throughput that felt like a gift. This is the part the productivity chart can never show you, because the productivity chart is measuring the wrong end of the pipe.
Faros AI's June 2026 “Acceleration Whiplash” study (~22,000 developers) and Opsera's 2026 AI Coding Impact Report, as reported: AI lifted throughput ~66% and cut time-to-PR up to 58%, but AI-written PRs wait ~4.6× longer in review (+91% review time), ship ~15–18% more security vulnerabilities, and incidents rise faster than throughput — while senior engineers capture ~5× the gains of juniors. Lens: Kevin Kelly, Out of Control, Law 7 — a complex adaptive system punishes single-metric optimization; the cost you stop watching doesn't vanish, it relocates. Industry reports, self-selected and correlational, not proof AI causes the incidents; figures as reported.
Out of Control: you can't yank one lever
Kevin Kelly wrote Out of Control in 1994 about exactly this failure mode, long before anyone typed a prompt. His subject was complex adaptive systems (economies, ecosystems, hives, code bases), and his warning was blunt: you don't govern a living system by grabbing one variable and pulling on it. Kelly's seventh law of making something from nothing is a single line — seek many goals, not one optimum — and the anti-pattern he names for it is optimizing a single KPI. A software organization is a vivisystem, not a clockwork. Yank the throughput lever and the system doesn't obediently go faster; it redistributes, routing the load you squeezed out of writing straight into reviewing, debugging, and firefighting. The metric moved. The system merely rebalanced around your grip.
The gap that widens
Here is the second-order effect that should worry you most, because it compounds. Opsera found senior engineers capturing nearly 5× the productivity gain that juniors got from the same tools. Read that slowly. The tool sold as a great leveler is tilting the field. Seniors have the judgment to sense when the AI is confidently wrong, so their throughput turns into shipped value; juniors get the same 66% more output and a thinner filter for catching the 15–18% that is quietly broken. A complex system rewards whoever already reads it best. Hand two people identical leverage and the experienced one pulls further ahead — so the divergence isn't only cost versus throughput, it's person versus person, and it keeps growing.
The honest caveat
Plant your feet before you quote these numbers at your standup. These are industry reports, not controlled experiments: self-selected teams, correlational data, vendors with a stake in the story. "Incidents rose while AI adoption rose" is not "AI caused the incidents"; a hundred things moved in 2026. The 66% and the 91% are real measurements of a real correlation, and they are not proof of a mechanism. But the pattern is too clean to wave off, and Kelly's is a structural argument that doesn't need the causation nailed down: when you optimize one number inside a system you don't fully control, the cost has to go somewhere, and it will pick the somewhere you aren't watching. You don't need a randomized trial to believe in conservation of trouble.
What this means for you
So stop celebrating the front of the pipe. Throughput up 66% is not a win; it's a number that tells you nothing until you know where the displaced cost landed. Go instrument the back end — review latency, escaped-defect rate, incident frequency, the security backlog — and watch those move together, because a complex system only tells the truth when you measure it whole. If your review queue and incident rate are climbing faster than your throughput, you haven't sped up. You've taken a loan, and Faros just showed you the interest rate. The fix isn't to rip out the AI. It's to stop pulling one lever and calling it flight. Fund the review capacity, the tests, and the senior judgment that turns cheap output into shipped value, all the boring downstream the throughput chart ignores. Give the acceleration somewhere to land, or the whiplash is just physics.
You optimized throughput and the cost didn't vanish — it moved to the queue you weren't watching. The cheap velocity wasn't bought. It was borrowed, and the incidents are the interest.
Instrument the back of the pipe, not the front: a complex system only tells the truth when you measure it whole.
Source: Kevin Kelly, Out of Control (the seventh law of making something from nothing — seek many goals, not one optimum; a complex adaptive system punishes single-KPI optimization, and the cost you stop watching doesn't disappear, it relocates). News pegs: Faros AI's June 2026 "Acceleration Whiplash" report, a study of ~22,000 developers finding AI lifted throughput ~66% while incidents rose faster than throughput; and Opsera's 2026 AI Coding Impact Report (time-to-PR cut up to 58%, AI PRs waiting ~4.6× longer in review with ~15–18% more security vulnerabilities, and senior engineers capturing ~5× the gains of juniors; high-adoption teams completing ~21% more tasks and merging ~98% more PRs as review time rose ~91%). These are industry reports — self-selected and correlational, not proof AI causes the incidents. Popular-science interpretation; figures are as reported.