Claim
AI in knowledge work raises the quality floor disproportionately. Human effort shifts from structural drafting to line-level rhetorical refinement, and the bottom of the quality distribution improves faster than the top.
Mechanism
A closed-loop taste flywheel drives the result. AI generates a full draft from seed ideas. The human reads and applies line-level verbal feedback, which the system uses to learn style preferences across sessions. Draft churn drops from 47 to 3 across 200 pieces because the AI now generates on-style first attempts. But line-level edits stay constant at a median of 136 per piece because taste-driven refinement cannot be delegated. The 10th-percentile quality score rises from 2.59 to 3.81 while the 90th-percentile improves by a smaller margin. AI eliminates the worst output by removing structural drafting failures. It cannot eliminate taste-driven variation at the top.
Conditions
Holds when: the human operator has a defined, articulable style that can be communicated through iterative feedback. Works best in high-output contexts where draft volume previously bottlenecked quality iteration.
Fails when: the domain requires original research or primary data that AI cannot generate. Also fails when the human operator lacks a consistent style to converge toward. Floor improvement disappears if baseline quality was already high.
Evidence
Tunguz measured his own output across 200 pieces. Draft churn fell from 47 to 3. Line-level edits remained constant at a median of 136 per piece. Quality scores on a 1-5 scale show the 10th percentile rising from 2.59 to 3.81, a larger gain than the 90th percentile saw. He draws a parallel to chess: the number of grandmasters tripled since engine-assisted training became standard, while training hours held flat.
"AI does not shrink human effort. It raises the ceiling of what the same amount of work produces."
Signals
- Draft count per piece falls sharply while editing time stays constant.
- Weakest pieces improve more than best pieces when AI handles structural drafting.
- Style consistency across pieces tightens over successive feedback cycles.
- First-draft acceptance rate rises without a corresponding drop in editing depth.
Counter-evidence
Tunguz's measurement is self-reported and from a single author's workflow, not a controlled study. The taste flywheel depends on the author having a stable, learnable style. Writers whose style varies intentionally by audience or register may not see floor improvement. The model for line-level editing assumes verbal instruction as the primary feedback channel, which is atypical for most knowledge workers.
Cross-references
- AI fails as a ghostwriter because every user shares the same underlying voice; it succeeds as an editor that suppresses itself: Tunguz's earlier finding that AI is a poor ghostwriter because it lacks the author's voice, requiring the author to define voice through feedback.