The Quiet Skill of Reviewing AI Work (And Why It Is Now a Core Competence)
Generating output is nearly free, which means the scarce skill moved downstream. Reviewing well is a craft with rules, and most people are winging it.
When output was expensive, generation was the skill. Output is now nearly free, and the scarce competence has moved downstream to review: the ability to look at a competent-looking draft and find the three things wrong with it. Almost nobody trains this, and it shows.
The core of the skill is knowing what you cannot check by reading. Fluency fools the eye; a fluent paragraph feels correct. So review has to be verification-first: every claim that carries weight gets traced to a source, every number gets recomputed, every quotation gets checked against the original. Reading is for structure; checking is for truth. Confusing the two is how fluent nonsense ships.
The second rule is reviewing against intent, not against quality. A draft can be well-written and wrong for the job: the right essay answering the wrong question, the accurate summary omitting the thing your audience actually needs. I keep my original intent in writing before generating, and the review asks one question: does this do the thing I wrote down?
The third is the calibration of changes. Novice reviewers rewrite everything (becoming the author, with AI as an expensive typist) or nothing (shipping whatever arrived). The craft is in the middle: keep what is right, fix what is wrong, and know which is which. That judgment is domain expertise wearing an editor's hat, which is why AI has made domain expertise more valuable, not less.
Teams that figure this out quietly pull ahead of teams still debating whether AI writes well. The question is settled; it writes adequately. Whether the work is good now depends entirely on whoever reads it before it ships. That person is the quality of the system, and it might as well be trained.