← All guides

Drafted with AI? Read this

How to clean up AI-written fiction (the “AI tells” problem)

If you drafted with AI, your manuscript almost certainly carries habits that make pages read machine-made — and readers are getting quicker at spotting them. The fix is not a full rewrite. It's a targeted cleanup: know the tells, find where they cluster, and rewrite those lines in your own voice.

What an "AI tell" actually is

A tell is a rate, not a single instance. One em-dash is ordinary craft. One sentence fragment is often the most human thing on the page. What reads machine-made is the pile-up: the same constructions, the same rhythms, the same stock phrases recurring page after page at a density no human drafts at. That distinction matters, because the goal of cleanup is to bring the rates down — not to strip every dash and fragment out of your book and your voice with them.

The tells to hunt first

  • Em-dash pile-ups. Two per paragraph, every paragraph — the single most-noticed tell.
  • The "not X, but Y" reflex. "It wasn't anger, but something darker." Once you see the negation-flip, you see it everywhere.
  • Echoed phrases and rhythms. The same sentence shape, the same pet constructions, recurring at machine-regular intervals.
  • Stock phrases. Breath she didn't know she was holding, a smile that didn't reach his eyes, something shifted between them.
  • The theme said outright. A character stating the book's message in dialogue that exists only to state it.
  • Arcs that resolve too neatly. Every chapter tying off its own tension — nothing ragged, nothing left pulling the reader forward.

The manual cleanup order

  • Search before you read. The find function turns invisible rates into a visible count: search the dash, "not", "but", and your suspect phrases. The number tells you how big the job is.
  • Fix by cluster, not by book. Tells concentrate in the chapters you touched least. Do a focused pass where they're thickest instead of a vague pass everywhere.
  • Rewrite in your voice, not "correctly". Replacing a machine phrase with a generic phrase trades one flatness for another. Say the line the way you'd actually say it.
  • Read the fixed chapters aloud. Your ear catches residual machine-rhythm your eye forgives.

Getting a map instead of guessing

The hard part of doing this by hand is knowing where you stand — which habits, which chapters, how dense. Starport's AI-Tell Cleaner reads the whole manuscript and marks every tell it finds as a comment in your own Word file: the passage, why it reads machine-made, and what you might do about it. It reads by rate, not single instances; it applies only the mechanical fixes that can't touch your voice (every one logged); and it hands you a density report with a per-chapter heatmap so you can see exactly where the patterns cluster. The Deep Edit goes further — model-judged easing of overused punctuation and rewrite suggestions drafted into your margins, each screened so it can't introduce a tell — but nothing changes a word of your prose without your hand on it.

Once the tells are out, the remaining question is whether the prose is good — tight, rhythmic, yours. That's a line edit's job, and the Line Edit measures every change against a Style Sheet built from your own writing, which is exactly what a post-AI manuscript needs.

Find out where your manuscript actually stands — free.

The Manuscript Scorecard grades your full manuscript on nine craft metrics, benchmarks it against professionally-edited books, and tells you exactly what to fix first. Free, in minutes.

Get your free Manuscript Scorecard →

FAQ

Will readers really notice AI tells?

Increasingly, yes — reviewers now name them specifically, and the em-dash pile-up in particular has become widely recognized. The tells read as flat or samey even to readers who can't articulate why.

Do I have to disclose that I used AI?

Amazon KDP asks publishers to declare AI-generated content at upload, and store policies keep evolving — check the current requirements where you publish. Disclosure rules are about generation, not about using AI tools to edit your own writing.

Can a tool catch every tell?

No, and any tool that claims to is overselling. What a good pass gives you is the map — what it found, where it clusters, and why it was flagged — so your own read-through spends its time where it matters most.

Will cleaning up the tells flatten my voice?

Only if you over-correct. Em-dashes, fragments and the occasional bold construction are ordinary craft — the problem is rate, not existence. Fix the pile-ups, keep the voice.

More guides