Can AI Detectors Tell the Difference Between AI Editing and AI Writing?

No — AI detectors cannot reliably tell AI editing from AI writing, because they score the statistical fingerprint of the final text with no access to your process, so a heavy AI rewrite of your own ideas can flag exactly like a fully generated essay.

Do AI detectors flag AI editing, or only AI writing? The question matters because a lot of honest people live in the middle ground. You wrote the draft. Every idea is yours. Then you asked a tool to “clean it up” — and now a checker says 78% AI and you’re staring at the screen feeling robbed. Were you? The answer hinges on where editing ends and writing begins in a detector’s eyes. Let’s map the whole spectrum.

Key takeaways

• Detectors see text, not process. There is no “edited” signal separate from a “written” signal — only the statistics of the words that survived.
• The spectrum matters: spell-check leaves your fingerprint intact; full-sentence AI rewrites replace it with the model’s. Flags start where replacement starts.
• “Improve this paragraph” is generation wearing an editing hat. The ideas stay yours; the statistical authorship transfers to the model.
• You can’t prove “only edited” from the text itself. Version history and saved before/after drafts are what prove it.
• Policies distinguish editing from writing inconsistently — the safe assumptions are local, not universal.

Why detectors can’t see your process

A detector’s entire universe is the character string you paste in. It computes how predictable each word is, how sentence shapes vary, how the phrasing compares to patterns learned from human and machine corpora. Nothing in that computation has a slot for *how the text got that way*. Turnitin says this plainly in its own documentation: the tool identifies text likely produced by a language model — it makes no claim about intent or workflow.

So the question “was this written or edited by AI?” isn’t one a detector answers. It answers a proxy: *whose phrasing dominates the final draft?* And that proxy is exactly why the middle of the spectrum gets messy.

Do AI detectors flag AI editing? The spectrum, from safe to indistinguishable

Picture your workflow on a dial.

Level 1: Mechanical correction. Spelling, punctuation, a misplaced comma. Your words, your structures, your rhythm — all untouched. Detector impact: effectively none. This is true even of tools with AI branding; we tested the common case in does Grammarly trigger AI detectors?

Level 2: Suggested micro-edits. Word-level suggestions you accept one at a time — a tightened phrase, an article fixed, a verb strengthened. Your fingerprint dominates; the changes are scattered and small. Detector impact: minimal, because the passage-level statistics are still yours.

Level 3: Sentence-level rewriting. “Rewrite this sentence to be clearer.” Now entire sentences carry the model’s phrasing preferences — its rhythm, its word probabilities. A few of these in a long document dilute into the background. Do it to a third of your sentences and the document’s statistics start migrating toward the model’s.

Level 4: Wholesale “editing.” Paste the draft, say “improve this,” accept the output. Here’s the uncomfortable truth: this is generation. The model produced every sentence of the final text, conditioned on your draft the way it’s otherwise conditioned on a prompt. The paraphrasing research (Krishna et al.) makes the equivalence concrete — a rewrite pass produces text with the *rewriter’s* statistical signature, which is precisely why paraphrase models both evade some detectors and flag on others. Your ideas survived. Your authorship, in the only sense a detector measures, did not.

The line, then, isn’t about which tool you used or what you called the activity. It’s about who chose the words that remain. A useful self-test: highlight the sentences in your final draft whose exact wording you produced. If you can’t find many, neither can a detector.

The honest-user trap — and how to escape it

The cruel part of this design is that it punishes transparency-minded people asymmetrically. Someone who generated an essay outright and someone who wrote a draft at 11pm and asked a model to smooth it produce final texts a detector may score identically. The difference between them — the thing that actually matters ethically — lives entirely outside the text.

Which tells you exactly where your protection has to live: outside the text.

1. Save the before. Every time AI touches something high-stakes, keep the pre-edit draft. The pair — your rough version, the polished version — is the single clearest evidence “edited, not written” can have.
2. Draft where history accumulates. A document with hours of incremental version history is its own alibi. A document that appears fully formed is a question mark, whatever the truth.
3. Apply suggestions by hand. If you retype or adapt a suggestion rather than accepting output wholesale, you keep both control of the phrasing and honest standing to say the words are yours.
4. Disclose where disclosure is expected. A one-line note — “Grammarly for grammar; no generative drafting” — costs nothing and reframes any later score dispute. We’ve collected wording that works in AI disclosure statements: when and how to tell your professor.
5. Check before you’re checked. If you’ve used level-3-or-above editing, score your edited draft yourself before submission. Better to discover the statistical migration while you can still fix it — usually by re-writing the flagged sentences in your own phrasing — than in someone’s office.

Where policy draws the line (spoiler: everywhere)

One reason this question stays confusing: institutions haven’t converged. Some syllabi explicitly permit “grammar and spelling tools” while banning generative rewriting. Some journals require disclosure of any AI involvement including editing. Some workplaces don’t care about anything but the output. The same workflow that’s compliant in one venue is a violation in the next, and the detector — inconsistent itself, as Weber-Wulff’s testing documented across tools and text types — is a rickety referee for all of them.

So don’t ask “is AI editing safe?” in the abstract. Ask “what does *this* venue’s policy say, and can I show my process if asked?” Those two questions have answers. There’s more of our writing on detection to help with the first, and our plans if the checking loop becomes routine. The tools can’t see the difference between editing and writing — which means, when it matters, you have to be able to show it.

Frequently Asked Questions

Do AI detectors flag AI editing? They can, depending on how heavy the editing is. Detectors have no access to your process — they score the statistical fingerprint of the final text. Light, targeted fixes (spelling, punctuation, a reworded phrase) leave your fingerprint mostly intact and rarely flag. But a full AI rewrite replaces your phrasing with model phrasing sentence by sentence, and the result can score exactly like generated text, because statistically it largely is.

Where is the line between safe AI editing and flaggable AI editing? Think in terms of who chose the words in the final draft. Grammar and spell-check corrections keep your word choices; so does accepting a suggestion here and there. The risk climbs when a tool rewrites whole sentences and you accept the output wholesale — “improve this paragraph” requests are generation wearing an editing hat. A rough rule: if you couldn’t point to which words are yours anymore, a detector can’t either.

Will Grammarly or a spell checker make my writing flag as AI? Standard grammar and spelling corrections almost never move a detector score meaningfully — they fix errors without replacing your sentence structures or vocabulary patterns. The risk comes from the generative features many of these tools now bundle: full-sentence rewrites, tone changes, and “make it clearer” buttons. Use the red-underline corrections freely; treat the rewrite buttons as what they are, which is AI drafting.

Can I prove my text was only edited by AI, not written by it? Not from the text itself — that’s the whole point; the final text is all a detector or an accuser can see. Proof lives in your process: version history showing the draft evolving in your hands, the original pre-edit draft saved separately, notes and outlines, and prompt history if you used a tool. If your workflow involves AI editing on anything high-stakes, save the before-and-after pair every time. That habit converts an unwinnable argument into an easy one.

Is AI editing treated the same as AI writing under academic policies? Increasingly, policies distinguish them — but they draw the line in different places, and some don’t allow either. One syllabus may permit grammar tools while banning generative rewriting; another bans all AI involvement including editing; a third permits anything with disclosure. There is no universal default, and “I only used it to edit” is not a recognized defense if the policy says otherwise. Read the policy for each course or venue, and when in doubt, ask before you submit.