Why "AI-Polished" Writing Can Score 100% AI-Generated (It's Not a Bug)

Quick answer

A detector can label a document "100% AI" without that meaning every single word was machine-written. Detectors typically break a text into segments, score each segment, and then calculate an overall score based on how many segments read as AI. When every segment shows even a faint AI signal, the document-level score reads 100%, even when a human wrote the draft and only used AI to smooth a few sentences. In August 2026, Nature identified this pattern on its own news sites. Articles that were human-reported and only AI-edited for translation or polish were marked 100% AI by Pangram, one of the two most widely used detection tools. That's not the detector malfunctioning. It's doing exactly what it was built to do, at a resolution most readers don't expect.

What actually happened at Nature's own sites

Nature News Feature went live on 25 August 2026. In that issue Miryam Naddaf and Richard Van Noorden tested Pangram's detector against articles from the Africa and India regional sites of Nature Portfolio. A number of the flagged pieces were AI-assisted translations of other Nature articles. Other flagged pieces were short research summaries that the authors had drafted with AI help amid a later full human edit. Some flagged items were original reports, including interviews, where writers disclosed that AI was used only to transcribe, translate, or organize notes. The editorial guidance from Nature Portfolio states that both categories are permissible. The publisher allows the use of AI for polishing or refining language without requiring disclosure; disclosure is required only when AI performs more substantial copy-editing or writing.

Pangram's older model called several of these pieces wholly AI-generated. When the site editors reran the same articles after the company shipped a new version, Pangram 4, in July 2026, most of them downgraded from "100% AI" to "mixed human-AI." But two news reports that contained original interviews and reporting still came back labeled 100% AI under the new model too. Pangram co-founder Max Spero, quoted in the piece, acknowledged the company still has "a lot of room for improvement" on real-world cases where AI and human writing blend together rather than sitting in separate, easy-to-isolate blocks.

Why polish and full generation can land on the same top-line number

The mechanism behind this isn't a mystery once you see how the score gets built. Pangram divides a document into segments, decides whether each one reads as AI, human, or mixed, then rolls those segment-level calls into one overall percentage based on the share of segments flagged. A "100% AI" verdict means every segment was judged probably AI. It does not mean every word in every segment was machine-written, because a segment can contain a mix of human and AI text and still get called AI overall if the model's signal in that chunk is strong enough.

That single fact explains a case in the same Nature piece involving Elena Vicario, director of research integrity at the publisher Frontiers. Vicario wrote a guest post for The Scholarly Kitchen in June 2026 arguing that journals should let peer reviewers use AI responsibly. Her own byline discloses that "Claude AI was used as an editing tool to help polish the text in later stage drafting" and that "all content, argument, and ideas were human generated." When Nature ran the post through Pangram in early July, it scored 100% AI. After Pangram 4 shipped, the same article scored 96% AI. Vicario's actual process, human ideas, human argument, AI-assisted line editing, never changed. What changed was how finely the detector was slicing the text and how it was weighting the slices it found.

When Pangram processed documents, it sliced them into sections of 200 to 300 words. An article of about a thousand words was consequently broken into three or four sections. Even a single poorly behaved slice can push the article into "AI-generated" territory. Pangram 4 trimmed the slice size to 30 or 40 words. This adjustment moved several Nature Portfolio items from a 100% AI label to a mixed label, even though the prose remained untouched. Spero coined the word "jitter" for the effect where a simple movement of a sentence, a paragraph cut, or a typo fix can trigger a larger jump in the AI score than the edit itself would warrant, simply because the text falls into a new slice.

The lesson isn't "detection is broken." It's "read the breakdown, not just the headline number"

None of this means the detector failed. Independent testing cited in the same Nature piece backs up the opposite conclusion: Pangram and GPTZero both scored zero false positives on 495 human-written passages in a July 2026 study from the research group Epoch AI, and Nature's own reporting notes both tools "correctly flag solely human-written content as human almost all the time." We've covered that Epoch AI dataset in more depth, including where detectors do miss real AI text, in our piece on style-imitation detection. The tools are accurate at the thing they're actually measuring. The problem is that "100% AI" reads, to almost everyone outside the field, like "100% of this was written by a machine," and that is a different claim than the one the number is making.

A percentage score is a probability estimate, not a fraction of a document. Treating that number as a verdict on a person's honesty is mathematically unsupported. Our guide explains what a raw percentage actually estimates in AI detection scores, and our piece on why 99% accuracy still isn't proof covers the deeper problem of using a single score as proof against a specific writer. The segmentation story adds a third layer of complexity: even if the population-level math is sound and the score means what it claims, the way a tool carves up a document before scoring it can significantly alter the headline number, for reasons that have nothing to do with how much of the writing was actually AI-generated.

The practical fix, per Nature's own reporting, is the one both Pangram and GPTZero already build into their paid tools: a sentence- or segment-level breakdown that shows which specific parts of a document triggered the flag, instead of stopping at one aggregate number. GPTZero's chief technical officer Alex Cui told Nature that determining whether a text is fully AI-generated is the tool's core job, and that measuring the exact extent of partial AI assistance is a harder problem. That's an honest limit to work around, not ignore. If you edited a draft with AI and it comes back flagged, the aggregate score alone won't tell you why. The segment-level view will.

FAQ

If my writing was AI-polished, not AI-written, can it still score 100% AI?

In cases where a manuscript has been AI-polished but not AI-written, the result is sometimes a blanket AI label. The literature reviews from the Nature Portfolio and Scholarly Kitchen highlight many of these instances. The detection tool performs segment-by-segment analysis. If each segment exceeds the high AI cutoff, the overall result becomes a 100% AI verdict. The catch is that this can occur even when a human wrote the original ideas and the AI only edited the phrasing.

Does a lower version number or a smaller AI edit always produce a lower score?

At Pangram, co-founder Max Spero describes "jitter" as the impact of small edits, such as removing a sentence or correcting a typo, on the position of a paragraph across chunk boundaries. The resulting score can change by more than the edit itself might predict. Yet a score change after such a minor adjustment does not necessarily signal more or less AI in the content.

What should I actually do if my AI-assisted work gets flagged as 100% AI?

That one-line "100% AI" label is a blunt summary that leaves out the details. If you want the truth, ask for the segment-level breakdown, that map shows every flagged portion, making it simple to show what was AI-assisted and what was purely yours.

Check your own draft before you submit it, not after someone flags it

This kind of segment-level analysis is exactly what TheChecker.AI runs, built for exactly this situation: you used AI for light editing, not full drafting, and you want to see precisely where a detector's signal is coming from before it becomes someone else's question to you. Run a paragraph or a full draft through our free demo and get a sentence-by-sentence view, not just one number to defend.