How to review AI edits before they quietly damage your book

By Derek Murphy · Updated September 29, 2026

The dangerous AI edit is not the ridiculous one.

You will catch a sentence that suddenly gives the detective three hands. The harder mistakes look helpful. The grammar is cleaner. The sentence flows. The character now says something sensible and emotionally mature—which is exactly why it is wrong for a furious seventeen-year-old who has spent the last ten chapters refusing to apologize.

AI editing works best when you treat every result as a proposal. Some proposals are obvious wins. Some solve the wrong problem. A few change the book while making the prose look more professional.

You need a review system that catches those quiet failures before you approve an entire manuscript because the first five corrections were good.

First, identify what kind of edit you requested

Judge the result against the job.

• Proofreading should fix mechanical errors with minimal rewriting.
• Copyediting may improve clarity and consistency, but it should preserve meaning.
• Line editing can reshape sentences, rhythm and emphasis, so voice risk is higher.
• Developmental feedback should diagnose patterns and story problems rather than silently rewrite the book.

If a proofreading pass rewrites paragraphs, that is a failure even when you like some of the new prose. The tool crossed a boundary you did not ask it to cross.

Save the original and make comparison possible

Never overwrite the only copy.

Keep the original manuscript, the edited manuscript and a small review log. The log does not need to be complicated. Record the chapter, the kind of problem, and whether you accepted or rejected it.

For example:

After twenty decisions, patterns appear. If the tool keeps flattening contractions, you know what to search. If it keeps helping with repeated words, that may be a good reason to use it again.

Check meaning before style

Read every changed sentence and ask a blunt question: does this mean the same thing?

Small wording changes can reverse responsibility, certainty or chronology.

Illustrative example:

Original: Mara thought Daniel had taken the key.

Edited: Mara knew Daniel had taken the key.

The edited sentence sounds stronger. It also destroys uncertainty and may ruin a later reveal. Thought and knew are not style variants. They are different facts.

Watch words that control certainty and causation: might, believed, seemed, because, before, after, only, almost, still, again. They are easy to smooth away and often carry the logic of the scene.

Check facts the model could not know

AI editors can invent a correction when the source is ambiguous. Check:

• names and pronouns;
• ages and family relationships;
• dates and elapsed time;
• distances and travel time;
• medical, legal, scientific and historical claims;
• fictional rules and magic systems;
• who witnessed an event;
• what each character knows at that point in the story.

A fluent sentence can contain a new factual error. Grammar quality is not evidence that the content is true.

Protect character voice

Read the edited dialogue aloud.

Characters should not all use the same sentence length, vocabulary and emotional intelligence. A line editor may remove repetition and awkwardness that reveal age, background, stress or personality.

Look for:

• slang made formal;
• contractions removed or inserted;
• profanity softened;
• hesitation cleaned up;
• blunt characters becoming polite;
• indirect characters suddenly stating exactly what they feel;
• jokes explained;
• subtext converted into information.

The original line may be less elegant and more alive.

Watch for generic emotional language

Models often make emotion clearer by naming it. That is sometimes useful. Across a whole novel, it can create the same polished fog in every scene.

Be suspicious when an edit adds phrases about hearts pounding, breaths catching, jaws clenching, eyes widening, emotions washing over someone, or a character realizing the weight of a moment. None of these phrases is automatically forbidden. Repetition and predictability are the problem.

Ask whether the action belongs to this character in this situation. If the sentence could be pasted into fifty unrelated novels, it probably is not doing enough work.

Check what disappeared

Editors are trained to remove clutter. Some apparent clutter is setup.

An AI pass may delete:

• a repeated object that becomes a clue;
• an awkward pause that signals a lie;
• an apparently minor character reaction;
• a sentence that establishes the geography of a later action scene;
• a deliberate echo between the opening and ending;
• a contradiction that is meant to reveal an unreliable narrator.

Review deletions more carefully than substitutions. A replaced adjective is visible. A missing clue may not hurt until eight chapters later.

Check consistency across chapters

Chapter-by-chapter editing creates a special problem: each chapter may be locally clean and globally inconsistent.

Search the full edited manuscript for:

• every major character and place name;
• titles such as Doctor, Captain or Mom;
• hyphenated terms;
• numbers written as words versus numerals;
• recurring invented terms;
• capitalization of organizations, magic and technology;
• spelling variants;
• repeated descriptions of the same person or object.

Also compare the end of one chapter with the start of the next. A tool working in separate chunks may repeat an explanation or change the emotional temperature between scenes.

Review the right amount at a time

Do not approve a 90,000-word edit by scrolling until everything becomes gray.

Use batches. Review one or two chapters, then stop. Keep a short rejection list. Search the rest of the manuscript for the same kinds of changes before reading every line.

A useful review order is:

1. changed facts and meaning;
2. deleted material;
3. dialogue and voice;
4. continuity across chapter boundaries;
5. style improvements;
6. punctuation and formatting.

This order catches expensive mistakes while you are alert. Commas can wait.

Do not confuse “different” with “better”

Line editing produces alternatives. An alternative can be smoother without being more effective.

Ask what the original sentence was doing. Was it slowing the reader before a reveal? Was it harsh because the character was angry? Did the repetition create pressure? Was the image strange in a way that belonged to the narrator?

Keep an edit when it solves a problem you can name. Reject it when its only advantage is that it sounds more like edited prose.

Run a final damage check

Before you treat the edited file as the new master:

• Compare word count by chapter. A large unexplained drop needs investigation.
• Search for missing scene breaks and formatting.
• Check the first and last paragraph of every chapter.
• Read all revised dialogue in one important scene aloud.
• Search names, dates and invented terms.
• Open the downloaded file in the program you actually use.
• Keep the original beside it until publication.

If the tool offers a side-by-side original and edited view, use it. A proper tracked-changes export would make this easier; do not assume you have one until the downloaded document actually contains reviewable change markup.

The standard is not perfection

AI editing does not need to be perfect to be useful. It needs to find enough real problems to justify the time and cost of reviewing its suggestions without creating more damage than it removes.

That is measurable. Use a chapter containing known errors and deliberate style choices. Count what the tool caught, what it missed and what it changed incorrectly. Then choose the editing level that gives you the best tradeoff for this manuscript.

The machine can make thousands of decisions quickly. Your advantage is that you know which decisions matter.

Compare an original chapter with an edited version in Chapterra, then review the result before starting the whole book.