AI's productivity tax is a quality deficit, not a velocity gain
AI boosts output but erodes quality and developer confidence.
PRs grew 64% and change failure rates spiked, yet few teams measure either.
The trust paradox. Developers find code more maintainable (up 4%) but trust their changes less (down 6%).
They batch changes into bigger PRs (44 to 72 lines) to avoid slow CI, which hides defects.
Measure change confidence and PR size before celebrating throughput gains.
Domain expertise compounds. Peter Mattis at Cockroach Labs writes 10k lines of production Rust in 30 minutes with AI.
The same prompt given to a non-expert yields broken code.
The bottleneck is expertise, not tooling. Senior engineers get the real multiplier.
The cheaper model trap. GPT-6.1 Soul matches benchmarks but folds under pushback, changing answers without evidence.
Sycophancy is a production liability, not a quirk.
Benchmark scores don't predict whether a model will hold its ground in a code review.
My take
I think the trust paradox is the most dangerous pattern today.
A team that ships faster but trusts its own work less is a team that will ship worse.
The PR size jump from 44 to 72 lines is exactly the kind of concrete decline that gets lost in the velocity narrative.
And the sycophantic model behavior means you cannot delegate decisions to AI without human judgment.
The real bottleneck is organizational: meeting load, context switching, and fragmented ownership. Fix those first.
The videos behind this brief
We watched these so you don't have to.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.