May 2026 Update: Better Humanizer, OOD Eval, 100% QuillBot Pass

What shipped in the May 2026 release

This is a changelog, so let me lead with the release notes and save the reasoning for the sections below.

• The rewrite now restructures sentences and paragraphs instead of swapping words for synonyms. That is the single biggest change, and the rest of this post explains why it matters.
• A cleanup pass runs after the rewrite to catch residual academic phrasing the rewrite leaves behind, the formal adverbs and essay-closer constructions that read as machine-formal even after a good restructuring.
• The reference corpus was refreshed so that every example our engine draws on is itself prose that passes major detectors. No synthetic or unvetted text in the pool.
• Out-of-distribution evaluation was added to the release gate. We no longer ship a version that only tested well on the kinds of essays we tuned on.

The headline result is the QuillBot internal eval moving from 16/30 to 30/30. That is not a marketing number we picked. It is the count of passages, out of thirty, that came back clean on QuillBot's own AI Detector after humanization. The rest of this post is about how to read that number honestly and what changed under it.

The structure-over-synonyms rewrite

Most humanizers paraphrase. They take a sentence, find the words a detector is most likely to flag, and replace them with synonyms. The sentence shape stays identical. The rhythm stays identical. The paragraph still opens and closes the way an AI model opens and closes paragraphs. Detectors caught up to synonym-swapping a long time ago, because the statistical fingerprint of machine writing lives in structure, not vocabulary.

This release rewrites at the structural level. Paragraphs get reorganized. Long uniform sentences get broken up and varied. The register comes down from formal-essay to something a person actually writes. The result is text that reads differently, not text that reads the same with different words. That distinction is the whole reason the eval number moved.

We will not publish the internal patterns the rewrite targets, because that is the part competitors would copy. The high-level claim is the honest one: this version restructures, it does not paraphrase, and the 16/30 to 30/30 jump is what restructuring buys you over synonym-swapping.

The cleanup pass for residual academic phrasing

A good rewrite still leaves residue. Certain phrasings survive almost any restructuring because they are how academic English is taught, not how any single sentence is built. Formal adverbs like considerably and substantially. Essay closers that begin with a summarizing pivot. Concept-definition openers of the X-describes-the-tendency-of-Y shape. A reader skims past them. A detector does not.

So after the rewrite, a cleanup pass catches that residue and resolves it. This is the layer that specifically helps writers whose first-draft register leans formal, which includes a lot of non-native English writers whose academic training pushes them toward exactly the vocabulary detectors penalize. If you have ever been flagged despite writing carefully, the formal-register residue is often why, and this pass is aimed straight at it.

The point of calling it out in a changelog is honesty about what improved. The rewrite does the heavy lifting. The cleanup pass closes the gap the rewrite leaves, and together they are why the matrix tightened this month.

The corpus refresh, and why the source pool matters

Our reference corpus is built from 2,590 real student essays. This release refreshed it so that every example the engine draws on is prose that genuinely passes major detectors, not merely human-written text that happens to be on file. The distinction matters more than it sounds. Human writing that a detector would flag is not a useful model for writing that a detector will pass.

Competitors who train on synthetic AI output, or who never vet their reference pool against detectors at all, do not have this baseline. They are teaching their tool to imitate the wrong target. Ours learns from text that has already cleared the bar it is trying to clear. That is the quiet reason a corpus-trained approach behaves differently from a paraphraser, and it is the reason the QuillBot result below is architectural rather than incidental. The full account of how the corpus is built and why 2,590 real essays beat a synthetic pool is in the 2,590-essay corpus deep dive.

What out-of-distribution evaluation actually means

Here is the scenario that makes this concrete. You wrote a psychology essay. The humanizer you used was tuned on persuasive essays, the kind everyone benchmarks against. It passed every test the vendor ran, because the vendor ran persuasive essays. Then your psychology essay got flagged anyway. The gap between the domain you wrote in and the domain the tool was tuned on is exactly what out-of-distribution testing closes.

Out-of-distribution, or OOD, evaluation means testing on domains the tool was never tuned on. Most tools quietly degrade the moment your subject leaves the common persuasive-essay norm, because that norm is all they were optimized for. A tool that passes 95% on persuasive essays and 40% on art history is not a 95% tool. It is a tool that fails most students writing about anything unusual.

For this release we evaluated across five domains the engine was not tuned on, including psychology, art history, environmental economics, and education theory. If your subject sits outside the persuasive-essay center of gravity, OOD evaluation is the only test that proves the tool covers you, and it is the test almost nobody publishes. The 30/30 QuillBot result was measured on this OOD set, not on home-turf passages.

How to read the 6-detector matrix honestly

So did it get better. Directly, here is the canonical May 2026 row, measured on the out-of-distribution set: 4% GPTZero / 8% Turnitin / 8% Originality AI / 6% Copyleaks / 30/30 pass on QuillBot's own detector / 3% ZeroGPT. Those percentages are AI-likelihood scores, so lower is better, and the QuillBot figure is a pass count out of thirty.

The reason we publish all six instead of one is that a single-detector pass is not a proxy for a cross-detector pass. Passing GPTZero tells you nothing about Turnitin, because they are built differently and weight different signals. Turnitin in particular runs a layered classifier with multiple signals, which is why it is usually the hardest cell in any honest matrix and why 8% there is a more meaningful number than any single-detector headline. If your school or your client uses a specific detector, read the cell for that detector and ignore the rest. The matrix exists so you can do exactly that, rather than trust one number and get surprised by another. The numbers above are public benchmarks anyone can reproduce by pasting humanized text into each detector.

The QuillBot self-fail diagnostic

The QuillBot result deserves its own section because of what it reveals, not because of the number alone. QuillBot ships two products, a humanizer and an AI Detector. If you have paid for QuillBot and used its humanizer, this is not a knock on your choice, it is a diagnostic worth understanding before you trust the output.

QuillBot Humanizer's own output returns approximately 95% AI on QuillBot's own detector, which we confirmed on an owner re-test. Read that again. The same vendor's humanizer and detector disagree about whether the humanizer worked, and the detector says it did not. That is an architectural self-fail. A synonym-swapping humanizer leaves the structural fingerprint intact, so the vendor's own detector still sees machine writing underneath the swapped words.

Our 30/30 pass on that same QuillBot classifier is the other side of the same coin. The vendor's detector is confirming an architectural difference between restructuring and paraphrasing, not rewarding a cleverer synonym list. That is the cleanest single piece of evidence that structure-over-synonyms is doing real work. For the full teardown of QuillBot's humanizer and detector together, see our QuillBot Humanizer review.

What this release fixes for your specific failure mode

Changelogs are easy to skim past, so here is the practical version organized by the situation you are probably in. If your draft starts as a ChatGPT prompt, the two-layer prompt workflow guide covers the drafting half before any of this applies.

If you passed one detector but failed another, that is the cross-detector gap, and the six-detector row is the only honest readout for it. Pick the cell for the detector that actually grades you and plan around that number.

If you humanized with QuillBot and then got flagged by QuillBot, you hit the self-fail described above, and a corpus-trained approach sidesteps it because it restructures rather than swaps.

If your essay is in an unusual domain and still gets flagged after humanization, you hit the out-of-distribution problem, and the five-domain OOD evaluation is the test that proves coverage outside the persuasive-essay norm. For the academic workflow specifically, what to do when Turnitin's layered classifier flags your work, see our Turnitin AI bypass guide.

If you write careful, formal English and keep getting flagged, the residual-phrasing cleanup pass is aimed directly at the formal adverbs and essay-closers that survive a rewrite and trip detectors. This is the improvement non-native English writers are most likely to feel.

What to expect, and where the free tier stops

A free account comes with four humanizations at 250 words each, and no card. That is genuinely enough to test the tool on a single passage and see the difference for yourself before committing to anything. I would rather you run it free first than take the matrix on faith.

It is worth being honest about where four free runs stop being enough. The 30/30 result above was measured across thirty out-of-distribution passages in one sitting. On the free tier you cannot reproduce that, because the account's four runs cover a single domain. For thorough pre-submission verification across a full essay, the Basic plan at $18 removes that ceiling. If the detector your institution uses is Turnitin specifically, the layered classifier is the hardest target in the matrix, and the Pro plan at $27 is tuned with that three-signal classifier in mind. None of this is a reason to skip the free test. It is a reason to know which ceiling you are about to hit before you hit it. For the full ranked cross-detector comparison across nine tools, see our best AI humanizer 2026 listicle.

Monthly cadence, and reading the right numbers

This post is updated monthly as the corpus, the rewrite, and the evaluation set evolve. That cadence is the point, not a footnote. Most humanizers never publish update notes at all, so you have no way to know whether the tool you used last term still behaves the way it did. The monthly changelog series is how you check.

It also means the canonical row in this post supersedes any figures you saw in an earlier month. Detector vendors update their classifiers, our corpus refreshes, and the calibration shifts. Before any submission that matters, read the current changelog row rather than a screenshot from a previous version. You can verify any of this yourself: the scanner that scored this update still runs on our /detect page, now under a trained model that decides the verdict, so paste any text and check it independently. Bookmark the changelog, trust the current row, and re-check before you submit.

By Fırat Mıhcı (ResearchGate). $0 affiliate stake.