How to review an AI-assisted draft

An AI-assisted draft can look finished before an editor has established what is true. Sentences connect smoothly, quotations fit neatly and recommendations sound reasonable. Those qualities make a draft easy to read. They do not establish the source of a number, the accuracy of a quotation or the usefulness of the advice.

A practical review separates three jobs: checking the evidence, fixing the structure and editing the voice. Trying to do all three sentence by sentence encourages premature polishing. It is frustrating to improve a paragraph’s rhythm and then discover that its central claim must be removed.

Use this process on a complete draft with its original brief and source material available. A short product announcement may need a modest version of each pass. A long research article will need a more detailed claim record and subject review.

Set up the review before changing sentences

Make a working copy and identify the person responsible for the final publication decision. Keep the original output available so the team can inspect where unsupported information entered the draft. Record which source packet and assignment brief produced it.

Read the article once without editing. Write down its promised outcome in a sentence. Then compare that outcome with the assignment. If a guide promised to help readers prepare an interview but mostly explains why interviews matter, that is a structural problem to address after the evidence pass.

Decide how to mark problems consistently. Three labels are enough for many teams: evidence needed, structure change and wording change. Reserve an explicit publication blocker for unresolved errors or missing approvals. A comment that says “check?” does not communicate the same urgency as “Do not publish until the product owner confirms this deadline.”

Pass one: inspect the claims and their sources

The Associated Press says generative AI output should be handled as unvetted material and judged against its editorial standards. Its generative AI standards describe the responsibilities of AP staff, not a universal policy for every publisher. The relevant lesson for a writing team is to review the evidence behind the output rather than treating fluent text as evidence itself.

Mark every externally checkable claim: dates, names, prices, product capabilities, statistics, quotations and descriptions of research. Recommendations also contain factual assumptions. “Send the file through the portal to preserve comments” assumes the portal supports those comments and that they survive the transfer.

Open the underlying source. Confirm that it actually supports the wording in the article. A source that says a feature is available to some accounts does not support an unconditional claim that every customer can use it. A survey of one profession does not establish the habits of all workers.

Pay special attention to numbers. Check units, time periods, denominators and the population being measured. An increase from 10 to 15 is five percentage points when those values are percentages; it is also a 50 percent relative increase. Which expression belongs in the article depends on the claim being made. Do the calculation independently and preserve enough context for the reader to understand it.

Separate evidence gaps from ordinary edits

Suppose an illustrative draft says: “The team’s new review process cut revisions by 40 percent and made every client happier.” The source packet contains a manager’s observation that revisions seemed to fall, but no counts or customer research.

Do not change this to “The process significantly reduced revisions and improved satisfaction.” The sentence still makes unsupported claims. Remove the invented measurement, accurately attribute the observation if it is relevant, and ask for actual evidence before describing customer satisfaction.

For quotations, compare the exact words with a recording, transcript or approved written response. Do not retain a polished quotation merely because it sounds consistent with someone’s view. If the quote cannot be verified, treat that as an evidence issue. Paraphrasing still requires a reliable account of what the person said.

NIST’s Generative AI Profile identifies confabulation among the risks of generative AI. That is a reason to verify plausible details, not a method for detecting them by style. A suspiciously generic sentence may be true; a precise and confident sentence may be false.

Maintain a compact claim record for complicated pieces. Include the claim, its source, the relevant passage, the review decision and any remaining question. This makes a later update easier because the next editor can see why a statement survived review.

Pass two: rebuild the reader’s route through the piece

Once unsupported sections are removed or held, review the structure. Read only the title, introduction, headings and conclusion. They should describe a coherent route toward the reader’s task. If the headings could fit almost any article about the topic, they probably need more specific work.

Place prerequisites before actions that depend on them. A guide to running customer interviews should explain how to select participants before presenting the interview schedule. A procedure should distinguish required steps from optional suggestions. If two sections answer the same question, combine them and use the space for a worked example or a missing decision.

Inspect each paragraph for its contribution. Does it add an instruction, evidence, a condition, an example or a useful explanation? A paragraph that simply repeats the introduction can usually go. Preserve necessary qualifications near the claims they limit so readers do not have to reach the final section to discover an exception.

Test the article with a realistic scenario. Imagine a reader following the advice with the information the page provides. Where would they stop to ask a colleague what to do next? That point may need an example, a definition or a clear boundary rather than another general tip.

Pass three: edit voice without changing meaning

Now work at sentence level. Replace vague promises with named actions. Resolve unclear pronouns. Use the team’s preferred terminology for the same object throughout the piece. Read passages aloud to catch sentences whose clauses compete for attention.

Preserve the strength of the evidence while simplifying. “May help” should not become “will” merely because the latter sounds cleaner. Do not turn a qualified research finding into a recommendation for every reader. An editor can shorten a sentence while retaining its conditions.

Watch for repeated introductions, decorative transitions and conclusions that restate every heading. Cut these when they delay useful information. Keep variation in sentence length, but do not make every sentence artificially short. Some ideas need a condition and an explanation together.

If you use an AI tool during this pass, constrain the request: revise selected wording, preserve numbers and qualifications, and report any proposed factual change. Compare the revised version with the approved facts. Another generated draft is another object requiring review.

Close the review on the final version

Run a final read after all changes are merged. Recheck links, names, headings, captions and calls to action. Confirm that a late rewrite did not reintroduce a removed claim or detach a caveat from the statement it qualifies. Check the rendered page where possible, because a table, link label or subheading can become confusing outside the editor.

Resolve the publication blockers explicitly. Record who approved the subject matter and which version received approval. If evidence remains missing, narrow the article to what can be supported or hold it for the answer. A deadline does not turn an unanswered question into a fact.

For your next draft, use one evidence pass, one structural pass and one voice pass before the final read. Keep the comments from each separate. The result should be a piece whose claims can be traced, whose order helps a reader act and whose language carries the intended meaning without unnecessary work.