Does Using AI to Edit Your Own Writing Trigger AI Detectors?

The short answer. On the evidence available, yes. A study submitted in August 2026 took human-written academic abstracts, applied only a light AI refinement of the kind many institutional policies permit, and found Pangram and GPTZero flagged them at 64 to 80 per cent. The same detectors flagged the unmodified versions at 9 to 15 per cent. A person wrote both. Only the polish changed. It covered abstracts and two detectors, so treat it as a strong signal rather than a settled rule.

Most writing about AI detection asks whether a machine-written essay can be caught. That question is well covered. The question underneath it is rarely asked and matters more to honest people: what happens to writing that a human produced, then tidied up with AI?

That is not a hypothetical use. It is what many institutional policies now permit by name. Use AI to check grammar, tighten an argument, improve clarity, but not to generate the substance. A great many students and academics are following that instruction exactly.

The evidence says a detector cannot tell the difference between that and cheating.

What the research actually found

The paper is Why AI Detection Fails for Academic Integrity by Jonathan A. Karr Jr, Grigorii Khvatskii, Ting Hua and Nitesh V. Chawla, submitted 6 August 2026. It tested published English academic abstracts across four domains, comparing older human-written work against contemporary work, and ran them through Pangram and GPTZero.

The relevant result is the light-edit condition. The researchers applied what they describe as a "refine abstract only" edit, explicitly framed as a proxy for guideline-compliant AI assistance. Not rewriting. Not generating. Refining.

Those lightly refined, human-authored abstracts were flagged at 64 to 80 per cent.

The same abstracts left unmodified were flagged at 9 to 15 per cent, and the paper notes non-STEM rates running far above STEM.

So across that corpus, a light refinement moved flag rates from single figures into most of the sample. Those are corpus rates rather than one person's odds, and the gap between the two conditions is the finding rather than any single number inside it. The authors' conclusion is blunt and worth repeating: detector scores should not serve as standalone misconduct evidence.

There is a second finding in the paper that completes the picture. When AI-generated text was run through commercial humanisation software, fewer than 4 per cent of those rewrites remained flagged. Put the two results side by side and the incentive structure looks inverted: on this material, lightly polished human writing was flagged far more often than machine writing that had been deliberately laundered.

Why does light editing move the score so much?

Because detectors do not look for authorship. They look for statistical regularity.

An AI model, asked to refine a paragraph, does what it is good at: it evens out sentence rhythm, replaces idiosyncratic word choices with more probable ones, and smooths the transitions. Those are exactly the properties detectors were trained to associate with machine text, because they are exactly the properties machine text has.

The human wrote every idea. The AI made the prose more predictable, and predictability is what the detector measures. A very light touch is enough to move it a long way. If you want the underlying mechanics in more depth, how AI detection actually works covers what these tools measure and what they cannot.

What our own testing says about the mixed case

We publish our benchmark results including the ones that do not flatter us, so here is the directly relevant number.

We tested passages that were half AI-written and half student-written. Our current model flagged 0 of 70 of them. Not a low rate. None.

We list that in our published figures as a stated blind spot, because that is what it is. Mixed-authorship text is the hardest case in detection and our model does not currently catch it. For comparison, on copy-pasted AI essays with no editing at all, the same model flags about 24 in 100.

Be careful what that number is and is not. It measures AI text our model failed to catch, not human text it wrongly flagged, so it is a miss rate rather than a false-positive rate. It does not tell you how our model treats a lightly polished human essay, because that is a different condition and we have not published a figure for it. What it does establish is that mixed authorship defeats our model completely on the case we tested, and the study above shows two other detectors failing on a different mixed case in the opposite direction. Nobody has this solved.

On genuinely human student work we wrongly flag 4 of 400 essays, with a 95 per cent range of 0.4 to 2.5 per cent. The full method and the hard cases are in how often AI detectors flag human writing.

What this means if you are marking work

Three practical consequences.

A score is not built to separate permitted assistance from prohibited authorship. If your policy allows AI for grammar and clarity but forbids AI-generated substance, the score does not carry the information you would need to tell those apart. It reflects how machine-like the prose reads, not how the text came to exist. The abstract study is the closest published evidence and it points one way.

A conscientious habit may raise the score. A student who runs a finished essay through a tool to catch errors before submitting is doing something reasonable. The study did not test students or essays, so nobody should quote a measured student risk. What it shows is that on the material it did test, a light AI refinement moved flag rates a long way upward.

Process evidence has become the actual answer, not the backup. Draft history, research notes, the ability to talk through an argument. JCQ guidance already requires students to declare AI use and retain evidence of their working, and that requirement is doing more work than any score can. We cover how that applies to coursework in what JCQ says about AI detection in coursework, and the wider false-positive problem in what teachers need to know about false positives.

What this means if you are the one writing

If you used AI within your institution's rules and your work came back flagged, there is now published research describing that pattern rather than it being a reflection on your honesty. The study is citable and specific. It covers academic abstracts run through two detectors, so cite it as that rather than as a study of students.

Keep your drafts. Version history in a document editor is the strongest single thing you can produce, because it shows work developing over time in a way no finished file can. Be able to explain your argument in your own words. If your policy requires you to declare AI assistance, declare it, and declare what kind. A declared grammar pass is a far better position than an undeclared one that a detector notices anyway.

If you want to see how your own writing scores before you submit, the scanner at Is It AI? shows which passages return a high AI-pattern score and what triggered each one. It is free and needs no account. It is a screening result and not a verdict. The same caution applies to every other tool discussed here.

The honest state of this

Detection tools are being asked to answer a question they were never built for. They were built to estimate whether text looks statistically machine-like. Institutions want to know whether a person cheated. Those questions came apart the moment AI assistance became a normal, permitted part of writing, and honest people are being caught in the gap between them.

Until a tool can distinguish a polished human argument from a generated one, and none currently can, the number on the screen should start a conversation and never end one.

Frequently asked questions

Does using Grammarly trigger AI detectors?
Nobody has published a figure for Grammarly specifically, and the 2026 study did not test it. What the mechanism suggests is that the more a tool rewrites rather than corrects, the more it replaces your phrasing with more probable phrasing, and probability is what detectors measure. A spelling fix and a generative rewrite are not the same intervention, and neither has a published flag rate attached to it.

What does a 30 per cent AI score actually mean?
Less than people assume, and it is not a proportion of the text. Our own score runs 0 to 100 and measures how strongly the writing matches patterns common in AI text; it is explicitly not the percentage of the text that is AI-written. Other tools present their numbers differently, so check what a given tool says its number means. None of them identify which tool was used, or establish that one was used at all.

If I only used AI to edit, can I prove it?
Not from the finished text, and that is the core problem for honest writers. You prove it with process: draft history, notes, timestamps, and being able to discuss your work. That evidence is stronger than any score and it is what a fair academic-integrity process should be asking for.

Are some detectors better at this than others?
They differ, and the published results point in different directions. The study found Pangram and GPTZero flagging lightly refined human abstracts at 64 to 80 per cent. Our own benchmark covers a different mixed case, passages that are half AI-written and half student-written, where our model caught 0 of 70. Those are separate experiments and should not be read as one comparison. Together they say that detection is least reliable on text of mixed origin.

---

Sources: Karr Jr, Khvatskii, Hua and Chawla, "Why AI Detection Fails for Academic Integrity", arXiv:2608.11256, submitted 6 August 2026. Is It AI? published benchmark figures, model v7, August 2026.