An AI edit can keep your numbers and still change your claim. In this fictional internal update, the limits of a pilot study are part of the result:
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.
A shorter rewrite might read:
Illustrative rewrite with errors
Our AI assistant cut response times by 25% across the support team.
The arithmetic is right: 9 is 25% less than 12. The sentence is wrong. It credits the assistant for a change the source cannot attribute, extends a small pilot to the whole team, and drops the distinction between median first response and response times generally.
Read each claim against its source
Keep the original beside the edit. For each changed sentence, ask what a reader would now believe. Then find the source passage that supports that belief. This catches changes that a search for names and numbers will miss, including shifts from “may” to “will” or from “some” to “all.”
In a 2020 study, Maynez and colleagues found unsupported content in summaries that neural models generated from source documents. The systems were earlier than today’s assistants, so their results do not give us a current editing error rate. They do show why a readable summary needs a separate check against its source.
FActScore, introduced by Min and colleagues in 2023, evaluates the support for individual factual claims rather than assigning one true-or-false verdict to an entire response. That research studied generated biographies. For everyday editing, the useful idea is to review claims individually: one sentence can mix a correct number with an unsupported explanation. The manual review here borrows that idea; it is not a validated implementation of FActScore.
Keep the detail the reader needs
In the pilot update, the routing change explains why the team cannot yet attribute the improvement. Removing it makes the result easier to sell and harder to interpret. A faithful edit can still tighten the wording:
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.
The revised version keeps the uncertainty where the reader can use it. If someone needs a claim about the assistant’s effect, the next step is to gather evidence that separates it from the routing change. Further wordsmithing cannot supply that evidence.
Ask for a review you can inspect
Give your assistant both versions and a specific review request. With Zero Slop installed, you can use this 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.
Inspect the cited passages yourself. An assistant can misread them, and a source can itself be wrong. Zero Slop’s local checks help flag writing patterns and changes to tracked details. Neither a lower writing score nor a passed detail check establishes that the final claim is true.
Once the claims hold, read the edit aloud. Restore a phrase if the rewrite has flattened your voice; cut a sentence if it repeats a point the reader already has. Accept the edit when it says what you can support, in language you would actually use.
Try Zero Slop on a draft or install the free skill in your existing assistant.