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

Quick answer

Not reliably, and 2026's best research says so plainly. A new peer-reviewed study out of Vrije Universiteit Brussel tested four detectors — Turnitin, Copyleaks, GPTZero, and Pangram — on papers that were fully AI-written, fully human, and a middle category that matters most in the real world: text that mixes AI-generated and human-written sections. The tools that catch a fully AI-written document almost every time can miss more than half of a mixed document, or, in the opposite failure, label a lightly AI-polished human draft as "100% AI." A parallel Nature investigation published the same month found the same pattern outside the lab: a research-integrity blogger's post, which she says was human-drafted and only AI-polished, scored 100% AI on one detector and 96% on another. The lesson isn't that detection doesn't work. It's that "AI or human" is the wrong question for most real writing, and a single headline percentage can't answer the question that actually matters: how much of this did AI write, and how much did it just help with?

Why "mixed" is now the normal case, not the edge case

Most people using AI now don't fit either box. Someone drafts a paragraph, then asks AI to tighten the phrasing. Someone writes an outline by hand and lets AI fill in a section. Someone transcribes an interview with AI help, then edits the result themselves. Each of those produces a document with AI fingerprints on part of the text — not none of it, and not all of it. Yet most of the public conversation about AI detectors still treats the question as a binary: either a human wrote it, or AI did.

The Vrije Universiteit Brussel study, published in the open-access International Journal for Educational Integrity in June 2026, built exactly this scenario to test it properly. Researchers Marijke Van Vlasselaer, Filip Van Droogenbroeck, and Bram Spruyt created 160 academic papers across four categories with a known "ground truth" percentage of AI content in each: 40 fully human-written papers (from before ChatGPT existed, so there was no chance of AI contamination), 40 fully AI-generated papers, 40 "hybrid" papers built by swapping a section of a real human paper for an AI-written one, and 40 "humanised" papers — the same hybrid text run through a second AI pass designed to sound more natural and evade detection.

What the numbers actually showed

On pure human text, none of the four tools misfired. Every one of them correctly cleared 100% of the genuinely human-written papers, and only GPTZero showed any measurable false-positive tendency, even that slight. Earlier detector research found much higher false-positive rates against human writing, so this counts as real progress. The study's authors call the near-zero rate here "an important and even perhaps reassuring finding" — false accusations against innocent writers have been one of the field's biggest ethical problems.

When a full AI draft is run through detection software, the reliability becomes uneven. Turnitin flagged 100% of those papers as human. Copyleaks, in a similar vein, misidentified 75% of the same content. GPTZero slipped past over 70% of the passages, only recognizing parts of what followed. Pangram, on the other hand, outright spotted AI use in 65% of cases. If partial credit is counted, Pangram's accuracy climbs to about 97.5% on the remaining material. All four tools — Turnitin, Copyleaks, GPTZero, and Pangram — were tested against the same set of AI-written papers.

The hybrid and humanised categories — the ones that actually resemble how most AI-assisted writing looks — are where the difference became stark. Turnitin's accuracy on hybrid text was 60%, and it dropped to 50% once that same AI-written section was "humanised" to sound more natural. Copyleaks scored 30% on hybrid text and 22.5% on humanised text. GPTZero scored a flat 0% on hybrid papers and correctly caught only one of 40 humanised papers (2.5%). Pangram was again the exception, correctly identifying AI content in 37 of 40 cases in both categories — a result the researchers partly attribute to its "mirror" training method, which trains specifically on AI text built to mimic a real human sample rather than generic AI output.

The researchers then pointed their best-performing tool at 1,163 real master's theses submitted at their own university, with no way to know in advance which ones used AI. Pangram flagged 529 of them — 45.5% — as containing some AI-generated content. But the detail that matters most for anyone worried about a false accusation is the how much: the median flagged thesis registered around 30% AI content. Cases above 80% were rare. Most flagged work wasn't AI-written; it was AI-assisted, in a document that was mostly the student's own.

Why a "100%" score doesn't always mean what it sounds like

Nature's news feature, published in August 2026, ran into the same problem from a different direction — following real disputes rather than a controlled study. Elena Vicario, a research-integrity director at the academic publisher Frontiers, wrote a guest post about AI in peer review. She says the ideas and first draft were entirely her own, and that she used AI only to polish the language afterward — "which is simply best practice," in her words. When Nature ran the piece through Pangram, it scored 100% AI. After the detector's newer model was released, that dropped to 96%. GPTZero, tested on the same text, returned 100% confidence that it was AI-generated.

The detector's own team explained why. Pangram's co-founder, Max Spero, told Nature that a "100% AI" score doesn't mean every word was machine-written — it means the tool divided the text into segments and judged each segment as probably AI, even when a segment contains a mix of human and AI writing. A paragraph that reads as clearly human in isolation can score as AI once it's combined with the surrounding text into one scored chunk, a quirk researchers in the piece called "jitter": small edits to a document can shift the score by a surprisingly large amount, because the boundary between segments moves. Pangram's own technical documentation found that when AI was used to heavily revise a human-written essay, the tool still called the result fully human 41% of the time — the same problem in reverse.

What this means if you used AI to edit, not to write

None of this means the tools are broken — Nature's reporting and the VUB study both point to genuine, measurable progress: false positives against honestly human-written text are now rare across most detectors, a real change from a couple of years back. What still hasn't caught up is the segment-and-score model's ability to describe degree, how much AI was involved, once a document isn't cleanly one thing or the other. Tim Requarth studies science communication at NYU Langone. He told Nature he doesn't put much faith in exact percentages once AI editing enters, especially down to individual sentences — exactly the situation for anyone who drafted something themselves and only used AI to tighten phrasing, fix grammar, or reorganize a section. A "high AI" score there is a real possibility, not proof the tool misjudged your process, just a limit on what one number can distinguish. Here's what's worth knowing before staring at that number too long.

A score is a starting point, not a verdict. Ours works the same way — see our detection-score guide for the details. Segment boundaries can flip a score, too. A small edit to your document changing the result is the "jitter" problem above, not proof your writing itself changed. And the all-or-nothing framing doesn't match how people actually write with AI now. The real question is how much assistance, and for what part — which is exactly why we publish our own methodology on the accuracy page instead of asking anyone to trust one number cold.

Feed a fully human draft into our detector, or a heavily AI-edited one, or something in between — you'll see the same kind of segment breakdown these studies describe, never just one blended percentage.

FAQ

What's the difference between "hybrid" and "humanised" AI text — and does a "100% AI" score mean nothing I wrote survived? Hybrid text mixes AI-written and human-written sections as-is. Humanised text pushes an AI section further, through another AI pass or manual editing, aimed at sounding more natural and slipping past detection — both are harder for most detectors to catch than a fully AI-written document, per the VUB study. And no, a "100% AI" score doesn't necessarily mean a detector thinks you wrote nothing. Pangram and some other tools split a document into segments and score the whole thing by what share look AI-written, so a document can hit 100% even when individual segments are human-edited or human-written, as long as every scored chunk read as probably AI. Check the segment breakdown, not just the headline number, whenever a detector gives you one.

If detectors are getting better at catching fully AI-written text, why do false accusations still happen? Most involve exactly the mixed-authorship pattern above, not clean AI-written text — substantially human writing that's AI-assisted, run through a tool built for a binary "AI or human" call. The VUB study found genuinely low false-positive rates on cleanly human text. What's unsolved is describing partial AI involvement, not confusing human writing for AI.

Check your own mixed-authorship draft before anyone else does

Whether you edited with AI, drafted with AI and rewrote by hand, or you're just curious what a segment-level breakdown looks like, run it through our AI detector and see the granular result the research above is asking the field to take seriously.