Grammarly Authorship: what it tracks, what the laundering test found, and what got fixed

The short version

Grammarly Authorship is a writing-provenance feature: instead of guessing whether finished text looks AI-generated, it watches the document while you work and labels how each part arrived — typed, pasted, AI-generated. That is the same basic bet Authmarks makes, so we follow it closely, and this article treats it the way we would want our own product treated: by what the vendor publishes, what independent testing found, and what happened next.

The one-paragraph history: Authorship launched in August 2024. In November 2025, Jonathan Bailey of Plagiarism Today showed that AI text run through Grammarly's own "Humanize" feature came out with a top-line label of "75% Typed by a Human." Grammarly shipped a labeling fix about a month later. Both halves of that story — the flaw and the fast fix — are worth reading closely, because they map the real difficulties of any process record, including ours.

What Authorship is, in Grammarly's words

Grammarly announced Authorship on August 14, 2024, as a way for writers to "demonstrate which parts of a document were human-created, AI-generated, or AI-edited." The announcement describes tracking across the apps where Grammarly already runs and sorting text into categories the company lists as "human-typed, AI-generated, modified with generative AI, pasted from a known or unknown source, and edited by Grammarly or a native spell checker." It rolled out first as a beta in Google Docs, with Microsoft Word to follow.

As of this writing, Grammarly's product page says Authorship "works in Google Docs, Microsoft Word, and the Grammarly Editor," and that it is "available in Grammarly Free and Grammarly Pro" — so the core capability is free, which we think is the right call (it is ours too).

Three design decisions stand out in Grammarly's own description:

• It is opt-in. Grammarly states Authorship "will not track text sources without your consent." Recording a writing process without consent would be surveillance; every serious entrant in this category has landed on consent-first.
• The report belongs to the writer. "Your school cannot see your Authorship Report if you don't share it." The record is an instrument the writer chooses to present, not a feed the institution watches.
• It is explicitly not a detector. Grammarly's page says plainly that "Authorship is not AI detection." It does not judge finished text; it reports what it observed while the text was produced.

If those three points sound familiar, it is because they describe the shape of writing provenance in general: consent-gated observation of the process, producing a record the writer controls. On the shape of the product, Grammarly and Authmarks agree almost completely.

The laundering test: November 2025

On November 6, 2025, Plagiarism Today published "How Grammarly Launders AI-Generated Content." The test was simple. Bailey generated text with Google Gemini, pasted it into Google Docs with Authorship active, and then ran Grammarly's own "Humanize" feature over it. The resulting Authorship report gave the document a top-line label of "75% Typed by a Human."

The fine print told a different story: the detailed breakdown showed 0% actually typed, with 75% attributed to "Rephrased with Grammarly's AI." The information was in the report. But as Bailey put it, "the text was still flagged as green, and a casual observer could easily be fooled into thinking that it was mostly human-written." His summary of the failure mode:

"If you copy and paste content from another source, either AI or human-written, and run Grammarly's 'Humanize' tool over it, Grammarly Authorship will mark it as 'Typed by a Human.'"

Note what this was and wasn't. It was not a failure of observation — the system saw the paste and saw the AI rephrasing, and recorded both. It was a failure of labeling: the top-line summary, the only line most readers would ever look at, presented machine-laundered text under the most trusted label the product has. A provenance record that observes correctly but summarizes misleadingly fails at the last step, where it matters most.

The fix: December 2025

A month later the story got its second half. On December 11, 2025, Plagiarism Today reported that a Grammarly representative had contacted Bailey after the November article, "promising to address the issues" — and the fixes shipped. The misleading category was relabeled "AI-Generated or Rephrased with AI," and text run through Humanize now lands in the copied/generated bucket rather than "Typed by a Human."

Bailey's follow-up also lists what remains open: text that Authorship cannot attribute — for example, output from Google Docs' built-in Gemini features, or anything written before Authorship was switched on — shows up as "Unknown," and the report cannot say anything about it. He closes by asking for better guidance for the educators who are handed these reports.

We did not find any standalone public statement from Grammarly about the episode; the documented response is the outreach to Bailey and the shipped label change, both reported in the December follow-up.

Credit where due: a vendor that fixes a provenance-labeling flaw within a month of public disclosure is taking the category seriously.

What the episode teaches about any process record

We would be writing a shallower article if we pretended these lessons apply only to Grammarly. They apply to every system that claims to record how a document was written — ours included.

1. The top-line label is the product. Hearing officers, editors, and teachers read the summary, not the appendix. If the summary can say "typed by a human" about text no human typed, the detail underneath does not rescue it. This is why Authmarks labels are conservative by construction: text the witness did not observe is labeled untracked, never silently folded into a human total, and machine-involved text (pasted, AI-suggested, programmatic) is labeled as such at the moment it happens. The record's honesty budget is spent on the headline.

2. A provenance tool must account for its own AI. The sharpest irony in the November test is that the laundering tool and the provenance tool were the same product. Any system that both rewrites text and vouches for text must label its own rewrites as machine contribution, at full strength, with no softer category for the house brand.

3. "Unknown" is a feature, not a bug — if it is loud. Bailey's remaining criticism is that untracked text is a blind spot. It always will be, for every vendor: no system can retroactively witness writing that happened before it was turned on. The honest design choice is to make the blind spot impossible to miss. That is our position too — see what a writing record cannot prove, which is the list of claims we refuse to make for exactly this reason.

4. The record should be checkable outside the vendor. An Authorship report is viewed through Grammarly. A sealed Authmarks record is a C2PA-signed artifact that anyone can check — including offline, including by people who distrust us. When a labeling dispute happens (and the November test shows they happen), an independently verifiable record lets a third party inspect the evidence rather than take the vendor's summary on faith. Why sealing matters at all is covered in tamper-evident vs tamper-proof.

If you are choosing a provenance tool

Questions worth asking of any product in this category — ours included:

• What does the top-line label say about pasted-then-rewritten text?
• Does the vendor's own AI assistance get labeled as machine contribution?
• How loudly does the report mark untracked or unknown text?
• Can the record be verified independently of the vendor, and does the check still work if the vendor disappears?
• Is there a human attestation step — does the writer review and put their name to the record before it is sealed, or is the report generated silently?

Authmarks' answers: machine text is labeled machine text at capture time; untracked text is labeled untracked and never counted as human; every seal is preceded by the writer's own review and named attestation; and the sealed record is independently checkable, free, with no account. You can create a free account and test those claims against your own writing — which is, after all, exactly what Bailey did to Grammarly, and exactly the kind of scrutiny this category needs.