An approval on a pull request proves that someone clicked Approve. It has never proved that anyone read the code, and with AI writing more of it every month, that gap is now measurable. In April 2026, Faros AI published two years of telemetry from 22,000 developers: as teams moved from low to high AI adoption, pull requests merged without any review rose 31.3%, and incidents per pull request rose 242.7%. Since September 21 I have been building Deslopping, a GitHub check that holds a risky AI-written change until the people merging it show they understand it. This is why I built it, and why I refused to build it the obvious way.
The approve button stopped meaning anything
The pitch was ten times the output. The measured version is smaller, but real. In its 2025 report, drawn from more than 10,000 developers on 1,255 teams, Faros found that developers on teams with high AI adoption completed 21% more tasks and merged 98% more pull requests. Average pull request size grew 154% and review time rose 91%. Faros named the bottleneck plainly: human approval.
A year later the bottleneck had found its own way out. In the 2026 data, task throughput per developer is up 33.7% and the merge rate per developer is up 16.2%, while median time in review is up 441.5%. Reviews got bigger and slower, and more of them were skipped. Bugs per developer rose 54%, monthly incidents 57.9%.
The code itself needs more reading, not less. CodeRabbit compared 470 open-source pull requests and counted 10.83 issues in each AI co-authored one against 6.45 in each human-only one, with logic and correctness problems 75% more common. Engineers know this. In Stack Overflow's 2025 survey, 84% of respondents use or plan to use AI tools, yet 46% distrust their accuracy and only 33% trust it. The biggest single frustration, cited by 66%, is "AI solutions that are almost right, but not quite." We distrust the output and approve it anyway, because there is too much of it to read.
What the outages say, and what they do not
The incidents that made the news are messier than their headlines. In February 2026, The Register, following the Financial Times, reported a 13-hour disruption to AWS Cost Explorer in one of Amazon's two mainland China regions after engineers allowed its Kiro coding agent to make changes. Kiro reportedly chose to "delete and recreate the environment". Amazon called it user error, specifically misconfigured access controls, and not AI.
In March, Fortune reported, again citing the FT, that Amazon held a mandatory "deep dive" into a run of retail outages. An internal briefing described a "trend of incidents" with a "high blast radius" relating to "Gen-AI assisted changes", among other factors, and an email from senior vice president Dave Treadwell said new guardrails would include more senior engineers signing off AI-assisted changes made by junior and mid-level engineers. Amazon told Fortune that only one incident discussed was related to AI, that none involved AI-written code, and that no such sign-off is required. A year earlier, Replit's agent wiped a live database holding data on more than 1,200 executives during a declared code freeze, as Fortune reported. Replit's chief executive called it unacceptable.
None of these is a clean case of a rubber-stamped pull request, and I will not pretend otherwise. What they share is a change that moved faster than anyone's understanding of it. The fix reported at Amazon is telling: another approval, from someone more senior. If an approval proves only a click, a more senior click proves the same thing.
I did not want another tool that watches engineers
The obvious product here is surveillance: track who approves fastest, flag the rubber-stampers, send their manager a chart. I refused to build it, for two reasons.
The first is practical. Reading cannot be proven. Scroll depth, time on page and expanded diffs are all easy to fake once people know they are measured. Comprehension can be proven, and faking it costs more effort than reading the code.
The second is the one I care about. A tool that feels like a test gets worked around, and it deserves to be. So Deslopping turns the question around: it interrogates the AI's output, not the engineer. A wrong answer is never recorded as a person's failure. It is read as the AI producing code that was not clear enough to understand in the time available. Managers see how the AI performs by repository, module, task type and AI tool, and never see a person. Any aggregate built from fewer than five distinct reviewers in a month is suppressed, so filtering to a module only one engineer touches reveals nothing, and no drill-down, export or admin override opens up an individual's attempts.
That has a cost, and I accepted it on purpose: a lead using Deslopping cannot find out who is rubber-stamping. The only named record a lead sees is a break-glass merge, an emergency skip that someone chose to make, with a stated reason.
What a gated pull request looks like
Deslopping installs as a GitHub App and scores every pull request from 0 to 100 on five signals: how sensitive the touched code is, how much depends on it, whether existing behavior changes, how those files have behaved before, and how much logic moved. Copy edits, CSS, docs and lockfiles skip the questions entirely. A logic change on a sensitive path, such as auth, payments or a migration, is gated at any size.
For a gated change, the engine reads the riskiest hunks with their callers and the repository's AI rules files, and writes two or three questions anchored to specific lines, each with a rubric for what a correct answer must contain. Before anyone sees a question, the engine answers it itself and grades that answer. A question its own rubric fails is thrown away, because a wrong rubric would mark a reviewer wrong for being right. Answers are a sentence or two of free text, not multiple choice. Each reviewer gets their own questions, so answers cannot be passed along in chat, and the author, who prompted the AI, is asked something harder: what does this change not handle?
A miss shows which lines to re-read and asks again, differently. After three, a colleague from CODEOWNERS or the files' recent authors is brought in to walk through it with you. Every question has a dispute button, and a disputed question is voided, never counted as a miss. In a real emergency, one click merges without the questions, logs who and why, and leaves a follow-up due within 24 hours.
Then comes the part I care about most. The reviewer rates the AI's output in a few taps, on comprehensibility, intent fit, shortcut risk and pattern fit. Over months those ratings draw a map of where the AI's idea of your codebase diverges from your engineers' idea of it, which is exactly where your prompts and rules need work.
Building it with the thing it checks
I am not writing this from outside the AI wave. Deslopping was built with an AI coding agent doing most of the typing: since the first commit on September 21, 2026, more than half of its 280 commits name Claude as a co-author, and 203 pull requests have merged.
The agent was fast, and it was confidently wrong in exactly the ways this product exists to catch. It reported a sign-in page as working because the route returned 200, while the page it redirected to was a 404. It delivered an onboarding screen with a component, stories and a clean accessibility suite, and no route to reach it. One run of fixes, pull requests 71 to 83, came entirely from seams I found by clicking through a real pull request: an event nobody sent, a route nobody linked. Later, after the full test suite and a real walk-through had both passed on a change to the gate itself, a second agent told to review it against the written specification still found five blocking bugs. The tests carried the same misunderstanding as the code.
That is the lesson the product is built around. Green checks tell you the code agrees with itself. Only a person who understands the change can tell you it agrees with the problem.
The 30,000-foot read
AI is not leaving the editor, and it should not. Faros's own numbers show more work getting done. What is breaking is the assumption underneath code review: that an approval means a person understood the change. Volume broke it, and more approvers, or more senior ones, will not fix it, because a click is a click.
What you can control is what an approval has to stand on. Keep AI-written pull requests small enough to read. Before approving, ask yourself the question Deslopping asks authors: what does this change not handle? Measure the AI, not the people, so engineers have no reason to hide where it lets them down. And keep a record of where the AI keeps getting your codebase wrong, because that record is worth more than any prompt template.
Deslopping does this on every risky pull request. It is in private beta, Team is €12 per active reviewer per month, and every team starts with a demo. I built it so engineers stay the ones driving, with the AI doing the typing.
Sources: Faros AI, Ten takeaways from the AI Engineering Report 2026 (April 12, 2026; 22,000 developers, more than 4,000 teams; 31.3%, 242.7%, 441.5%, 57.9%, 54%, 33.7%, 16.2%); Faros AI, The AI Productivity Paradox Report 2025 (July 23, 2025; more than 10,000 developers, 1,255 teams; 21%, 98%, 154%, 91%, human approval as the bottleneck); CodeRabbit, State of AI vs Human Code Generation Report (December 17, 2025; 470 pull requests; 10.83 versus 6.45 issues, 75%); Stack Overflow 2025 Developer Survey, AI (84%, 46%, 33%, 66%); The Register (February 20, 2026; Kiro, 13 hours, Cost Explorer, Amazon's statement); Fortune, March 11, 2026 (Amazon's meeting, briefing, Treadwell email, Amazon's denial); Fortune, July 23, 2025 (Replit, more than 1,200 executives, code freeze, chief executive's response); Deslopping (how the product works, price, private beta). CAVEATS: Faros sells engineering analytics and CodeRabbit sells AI code review; their figures come from their own studies, not independent audits, and Faros compares low and high AI adoption, which shows association, not cause. CodeRabbit inferred AI co-authorship from signals in each pull request, and its sample is small. The Amazon accounts rest on Financial Times reporting from unnamed sources, which could not be opened from here, and Amazon disputes the AI attribution in both cases. The Replit details come from the founder's own posts as reported by Fortune. The figures about how Deslopping was built come from its own repository and build notes, and its description from the product specification; it is in private beta, so this article claims no customer results.