Proofreading When the Draft Is Already Clean

Anthropic shipped a cite-checking tool for legal teams this week whose job is comparing a document against its sources and flagging quotes that do not match. That is a telling thing to need to build, and it describes where proofreading has gone generally. The errors worth hunting are no longer on the surface.

What proofreading used to catch

A rushed document looked rushed. Typos clustered where attention had lapsed, sentences got tangled where the writer was unsure, and a reader skimming for errors was also, without meaning to, locating the parts of the document that deserved a second look.

That correlation did real work and nobody designed it. Surface quality and underlying care travelled together because they came from the same person on the same afternoon.

They have come apart. A document can now be flawless in construction and wrong in substance, with nothing on the page distinguishing the two, which is the same loss of cues described in quality assurance when AI writes the first draft.

The six categories to check

Check	How it fails	How to find it
Numbers and arithmetic	Plausible figure, wrong total	Search every digit, redo the sums
Names	A real name that is the wrong one	Check against the source record
Dates and deadlines	Internally consistent, externally wrong	Check against the calendar and contract
Commitments	Promises more than you offer	Find every will, shall and guarantee
Attributed claims	Source exists, does not say that	Open the source and read the passage
Omissions	Nothing on the page to notice	Check against a list, not by reading

Attributed claims are the category that embarrasses people publicly. A citation to a real source that does not support the sentence attached to it reads as perfectly normal, and the only way to catch it is to open the source. There is no shortcut and the check is quick once it is habitual.

Omissions are the costliest and the least checkable by reading. A generated document covers what the brief asked for, and what is missing is usually what nobody thought to mention: an exclusion, a dependency, a payment term, a caveat that would have changed the reader's decision.

Search, do not read

Reading a document end to end feels thorough and is the least efficient way to find these errors, because fluent prose gives a reader no reason to stop. Searching is better.

Run a pass for every digit in the document. A pass for proper nouns. A pass for will, shall, guarantee and within. Each takes under a minute on a short document and lands your attention precisely where the consequences are.

For omissions, keep a short list per document type of what must appear, and check the list rather than trusting that an absence will feel wrong. That is the same mechanism as a pre-flight check, and the design of one is in checklists that hold up.

Where AI helps with its own output

Consistency sweeps are the genuine use. Asking whether a name, a figure or a term is used identically throughout a long document is mechanical comparison work that software does better than a tired person, and inconsistency is frequently the visible edge of an underlying error.

Verification is where it stops. The model that produced a confident wrong number will confirm that number when asked, and a second model is somewhat better and still not a source. Checking a claim means opening the thing it came from.

The one exception is structured comparison against a document you supply, which is what a cite-checking tool does: not asking the model whether a claim is true, but asking it whether this passage supports that sentence. That is a comparison task with both sides present, and it is a different and much safer question.

Who should do it

Not the person who generated the draft, for the same reason writers have always needed editors, and more so now: having prompted it, they know what it was supposed to say, which makes them the least able to notice that it does not say it.

Fatigue matters more than it used to as well. An hour of reading competent, plausible, mostly-correct material degrades attention faster than an hour of obviously rough drafts, because nothing interrupts the flow. Shorter sessions catch more, and rotating who checks catches more again.

Reserve this properly for documents where being wrong has a consequence outside the room, and let internal drafts through with a lighter touch. Checking everything equally is how checking stops happening, which is the argument in the workslop problem. Where you are writing claims about your own work, the verification discipline is in writing case studies, and the tool landscape is in the best AI writing tools.

Frequently Asked Questions

Is proofreading still necessary if AI wrote the draft?

More necessary, and for different errors. Spelling, grammar and sentence construction arrive clean, which removes the surface problems and also removes the signals that used to tell a reader where to look. What remains is wrong numbers, invented sources, confident claims that nobody verified and quiet omissions, none of which look like mistakes on the page.

What should a proofreader check now?

Six categories, in order of consequence. Numbers and the arithmetic between them. Names of people, companies and products. Dates and deadlines. Anything stated as a commitment or guarantee. Any claim attributed to a source, checked against that source. And what is missing, which is the hardest and often the most expensive. Reading for flow is pleasant and is no longer where the risk sits.

Can AI proofread its own output?

It catches mechanical issues well and is unreliable on the errors that matter, because the same model that produced a confident wrong figure will confirm it. A second model is somewhat better and still not a substitute for checking against a source. Use it for consistency sweeps, such as whether a name is spelled the same way throughout, and keep verification with a person and the original document.

How do you proofread for things that are missing?

With a list rather than by reading. A generated document covers what it was asked to cover, and the gap is usually something the brief did not mention: an exclusion, a caveat, a dependency, a payment term. Keep a short checklist per document type of what must appear, and check against the list rather than hoping the absence will feel wrong while reading.

How long should this take?

Less time than reading the whole document, if you check categories rather than lines. Searching for every number, every name and every date in a five-page document takes a few minutes and catches most of what matters. A full read-through takes longer and reliably catches fewer of the errors that cost money, because fluent prose does not prompt the reader to stop.