GitClear: +25% output, +81% duplication — the maintainability tax
Sep 14 on Maintainability Gap (623M changes, 2023–2026): heavy AI users +25% vs own baseline; block duplication 40.3→73.0/MLOC; moved/refactor code 21%→3.8%.
Source: Your AI coding spend bought 25% more output. Duplication rose 81%.
Steve Fenton's Sep 14 New Stack piece digests GitClear's Maintainability Gap report (623 million analyzed changes, 2023–2026) into the invoice continuous generation forgets to print. Heavy AI users gained ~25% on their own prior velocity — real, and nowhere near 10× marketing. Those same teams already outpaced peers before AI; the tool amplified an existing distribution. Meanwhile block duplication rose 81% (40.3 → 73.0 per million changed lines), and moved code — the refactoring signature — collapsed from 21% of changed lines in 2022 to 3.8% in 2026. Pre-AI, developers chose refactoring over copy-paste ~2:1; now they are ~5× likelier to paste.
The tortoise framing is economic, not aesthetic. Early in the cost curve you get away with it; later rework steals feature time and "minor" bugs become economically unviable to fix. Fenton's point for budget season: if you measure LOC, PR count, or feature tickets, you miss the maintenance bill that compounds under always-on generation. Teams that already treated test automation and refactoring as the work (not a side quest under delivery pressure) are the ones still converting AI spend into durable throughput.
Why it matters for code review
Unverified overnight volume without a maintainability harness is the hare with a deferred tax. Review under sustained agent load has to catch duplication and missing refactors the same way it catches wrong behavior — or the factory's merge rate looks healthy while the codebase quietly becomes unreviewable. Pair continuous generation with continuous maintainability gates (dup/hotspot signals, small batches, refactor budget) or Cherny's verification stack will be grading an ever-worse substrate.