Updated · 4 minute read

How to check an AI edit.

Check an AI edit for changed facts, missing caveats, and unsupported claims. Use a worked example and review checklist before publishing.

By Manav Mishra

After reviewing dozens of AI edits, I learned to check the claims before the grammar. A rewrite can sound better while making a promise the source never made. This fictional internal update shows how it happens:

Source draft

During May’s pilot, 12 support agents used the assistant. Median first-response time fell from 12 to 9 minutes. Queue routing changed that month too, so we cannot yet separate the effects.

Illustrative rewrite with errors

Our AI assistant cut response times by 25% across the support team.

Nine is 25% below 12, so the calculation is correct. But the rewrite extends a result from 12 pilot participants to the whole support team and replaces median first-response time with response times generally.

It also credits the assistant for an improvement whose cause the pilot could not isolate. Queue routing changed during the same month; the source explicitly says we cannot separate the effects. Keeping the number did not keep the claim intact.

May pilot · 12 agents12 min9 minmedian first responseQueue routing also changed.This rewrite adds a cause:AI cut response times 25%.
A fictional example. The percentage is correct; the claim of causation is unsupported.

Read the source beside the edit

For every changed claim, ask what a reader would now believe and which passage in the source supports it. Check who is included and what was measured, then look for any new claim about the cause.

Watch for quiet shifts from “may” to “will,” from a pilot to a rollout, or from association to cause. A scan for names and numbers will miss those changes because the same details can appear in a different claim.

In a 2020 study, Maynez and colleagues found that fluent model-generated summaries could contain unsupported statements. That result explains why a readable edit still needs a source check. The study used older models and does not establish an error rate for today’s tools.

FActScore, introduced by Min and colleagues in 2023, evaluates a response one factual claim at a time. It was developed for generated biographies. The practical habit transfers to editing: split a sentence into the claims it asks you to accept, then check each one against the source.

Zero Slop uses that claim-by-claim discipline but is not a validated implementation of FActScore. The research supports the need for source checking; it does not validate this particular workflow.

What changed in the claim?fellcut12 pilot agentsthe support teammedian firstresponseresponse timescause addedgroup widenedmetric changed
Check who is included, what was measured, and whether the edit adds a causal explanation.

Do not edit away the caveat

The routing note explains why the team cannot claim causation. Deleting it would change the claim, even if the remaining sentence sounded cleaner. I would keep it in the edit:

Reviewed edit

In a May pilot with 12 support agents using the assistant, median first-response time fell from 12 to 9 minutes. Queue routing also changed, so the assistant’s contribution is still unclear.

Keep a record of the changed claims

For a longer edit, reviewing from memory is hard. Keep the source and rewrite side by side, then make a short record for each sentence that changes a factual claim. The pilot example would produce these entries:

Copy the supporting source passage beside each entry. If the rewrite adds a claim with no supporting passage, mark it for the author to verify or remove. A plausible explanation from the assistant does not supply the missing evidence.

Check instructions as carefully as reported results. An edit that drops an exception, changes who must approve a request, or moves a deadline can change what the reader does. Compare those details before polishing the wording.

Decide what you can publish

A source comparison checks whether the edit preserves the supplied account. To establish whether that account is accurate, check its underlying records or references too.

Give your assistant a better brief

Give the assistant both versions and name the changes it should look for. With Zero Slop installed, use this review prompt:

Use Zero Slop to compare this edit with the source. Flag changes in facts, scope, certainty, and causation. Quote the source passage for each proposed correction. Keep my voice. If the source is missing evidence, ask me for it before rewriting.

Read the source passages yourself after reviewing the flags. Zero Slop’s local checks flag stock writing and changes to tracked details. They cannot prove that the source itself is true.

Read the result aloud and restore anything that sounds unlike you. Cut repeated points, then check that the remaining sentences still say what you intended.

Try Zero Slop on a draft or install the free skill in your existing assistant.

← All posts