What happens to a code reviewer who sees their hundredth AI-generated pull request?
The intuition runs one way. Repeated exposure should produce expertise. A reviewer who has seen a hundred AI-generated PRs has a hundred data points about how AI code fails — the hallucinated imports, the plausible but wrong abstractions, the test coverage that only targets the happy path. Month over month, the gate should tighten.
A KDD 2026 study tracked 400 code reviewers over seven months — 11,429 reviews of AI-generated pull requests. The gate didn’t tighten. It opened.
Approval rose 14.5 percentage points — from 30.1% to 36.8% — with statistical significance beyond p<10−6. Inline comments dropped 22%. Review latency expanded 3.5 times: reviewers spent more time per PR but left fewer substantive comments. They looked longer and said less.
During the same seven months, approval rates for human-written code declined. Reviewers got stricter with human code and more permissive with AI code. Same reviewers. Same codebase. Opposite directions. The authors named it reflexive habituation under growing workload.
The Volume
This isn’t a research artifact. JetBrains’ 2026 developer ecosystem survey — 15,000+ developers — reports 47% of code is now fully agent-generated. Thirty-one percent of developers are “agentic coders” generating 84% of their output through agents. Go and JavaScript developers are at 54–55%.
The code stream that habituates reviewers is growing faster than reviewers can resist it.
Three Fixes That Miss
| Response | Assumes the problem is | Proposed fix |
|---|---|---|
| Better tooling | Effort | Risk scoring, trust-calibrated diffs, walkthrough modes |
| Better process | Organization | Checklists, reviewer rotation, team structure |
| Remove humans | Bottleneck | Agent ensembles reviewing agent code |
Each addresses a real problem. None addresses the mechanism the data identified.
The habituation finding is cognitive, not logistical. The same reviewer who tightens standards for human code loosens them for AI code — not because the tooling is bad, or the process is broken, or they’re too slow. The critical response declines because repeated exposure to plausible output suppresses it. This is not review fatigue from volume. It is habituation: a specific, measurable decline in evaluative scrutiny from repeated stimulus.
Better tooling helps if the problem is effort. Better process helps if the problem is organization. Removing humans helps if humans are the bottleneck. If the problem is that exposure degrades the evaluative function, then increasing exposure — through any of these channels — may accelerate the thing you’re trying to fix.
The Alternative
The “remove humans” response has its own failure mode. A recursive self-training study tested what happens when AI systems review AI-generated code. The result: a rubber-stamp regime. Acceptance rates rise while code correctness falls. The filter converges toward an identity operation. The authors proved it formally — Theorem 2.3: when the AI gate becomes constant on the generator’s output distribution, gated training equals ungated training.
Human reviewers habituate. AI reviewers rubber-stamp. The gate opens either way.
Sixty People
Trust-calibrated review tooling exists. A JetBrains participatory design study proposed a three-level review workflow — overview, file analysis, code snippet — with per-file risk scoring and security cages for high-risk changes. Sixty-three percent of participants expected it would reduce review effort while maintaining quality.
The study included 60 people.
The “3100 Opinions” study found that teams can set the direction — the expertise reviewers bring and how the team structures its process determines whether AI code gets waved through or genuinely evaluated. Habituation is not destiny.
But it is the default. And the question the data leaves open — that no study has run long enough to answer — is whether a reviewer who has habituated over seven months can reverse it. Whether trust-calibrated tooling deployed at production scale, not with 60 people but across the 47% of code that’s now agent-generated, can close a gate that exposure has opened.
Or whether the only thing that closes the gate is a failure large enough to reset the response.