EU Text Provenance Rules: Our Approach and What to Do Now

Our approach to EU text provenance rules is simple: provenance labels say where text came from, not who wrote it, and they do not change how you edit. If you publish into the EU, check the primary source at openai.com/index/eu-text-provenance, keep your drafts, and fix the writing you can actually control. The label is metadata. Your manuscript is still yours.

What actually changed this week?

A provenance layer now attaches to text produced through certain AI systems, and the people who feel it first are the ones shipping content into EU markets: marketing teams, agencies with EU clients, researchers submitting to EU-hosted journals, and anyone whose CMS pipes through a model API. The change is about traceability of origin, not about scoring your prose. Nothing in it measures whether a paragraph reads machine-made.

That distinction matters, because most of the panicked email threads I have seen confuse two different things:

Signal	What it reports	What it does not prove
Provenance metadata	Which system generated or processed a passage	Who wrote the ideas, or whether a human edited it
AI detector score	Statistical signal in the wording	Authorship, intent, or misconduct
Similarity percentage	Overlap with a reference corpus	Whether the overlap is quoted, cited, or accidental

A model score, a share of qualifying prose, and a similarity percentage are not interchangeable. None of the three is a verdict on authorship. If a client, editor or committee treats any one of them as proof, that is a process problem, not a writing problem.

Who gets hit first, and who does not?

The practical exposure is narrower than the headlines suggest. You are affected if your text passes through a system that records provenance and you publish into the EU. You are largely unaffected if you write in a plain text editor, paste into a CMS that does not call a model, or publish outside EU scope.

The people who get caught in the middle are the ones who used AI as a drafting aid and then edited heavily by hand. Their final text may carry a provenance record from the drafting stage while reading nothing like the model’s output. That is the awkward case, and it is worth naming honestly: a provenance label can be technically accurate about the tool and completely uninformative about the writing.

Here is the sequence that keeps you out of trouble:

1. Export the primary source page to PDF with the date visible.
2. Check whether your publishing stack sends text through a model API. Ask your platform owner, do not guess.
3. Keep your own version history: first draft, model-assisted draft, final edit.
4. Note in your editorial log which sections were model-assisted and which were not.
5. If you must disclose, disclose the tool and the stage, not a confession about authorship.

If the label on your text worries you more than the writing does, run the draft through the AI detector first. Sentence-level scores show which lines carry the signal, so you edit the lines that need it instead of rewriting a whole document blind.

Does a provenance label mean a detector will flag me?

No, and this is the objection I hear most: “if the label is there, the detector will match it.” Provenance and detection run on different machinery. Provenance is a record attached to a generation event. Detection is a statistical read of the wording in front of it.

We tested this on our own material. On 2026-09-28 we scored two corpora with the DeAIze detector (local ensemble, lite tier) at a 30% flag threshold. A reference sample of 709 passages from DeAIze guides not marked generated returned an AI score mean of 26%, median 25%, range 0-69%, with 239 of 709 passages at or above the threshold. Those are passage counts, not counts of independent documents, and human authorship was not independently verified. Unedited model output, 12 passages from fixed prompts, returned a mean of 33%, median 34%, range 4-46%.

Read those two ranges next to each other. They overlap. A passage at the low end of the model-output range sits below a passage at the high end of the reference range. That is the whole story of why a single number cannot settle an authorship question, and why a provenance label plus a detector score still does not add up to proof.

The evergreen fix: edit the sentences that carry the signal

Whatever the label says, the editing work is the same, and it is the work described in our pillar guide on humanizing AI text. You are not trying to beat a tool. You are trying to make the text sound like a person with something specific to say.

Start with the passages a sentence-level detector highlights. In our own sample, the AI-tell frequency ran at 1 per 100 words in the reference corpus and 1 per 100 words in the unedited model output, so frequency alone will not separate good prose from generated prose. What separates them is placement: generated text tends to spread the same habits evenly, while human drafts cluster their quirks.

A worked illustration, not a real client document:

• Before: “It is important to consider the multifaceted nature of provenance requirements in order to ensure comprehensive compliance across all relevant stakeholders.”
• After: “Provenance rules apply to three groups: publishers, platforms, and anyone piping text through a model API. Check which one you are.”

The second version names a number, names the groups, and drops the padding. It also changes the claim, which is the part people skip. Any rewrite that alters a quantity, a citation, a negation or a qualifier has damaged the document, even if the score drops. Compare every revised sentence against the original and verify the facts yourself. Semantic similarity is a review aid, not proof that meaning survived.

When this is the wrong choice

Do not buy anything for this if you are in one of these situations.

• You are in an authorship dispute. A rewrite tool is not evidence, and editing the disputed text can destroy the record you need. Preserve the original, keep timestamps, and follow your institution’s process.
• Your policy bans AI assistance outright. Then the fix is not a rewrite; it is a conversation with your supervisor about what you actually did.
• You never used a model and your text is clean. Provenance rules are not a reason to run your work through a paraphraser.
• You need a Turnitin number. Nobody can promise you one. DeAIze does not provide an official Turnitin readout or a Turnitin score estimator, and any tool that claims otherwise is selling you a guess.

If your situation is a committee or a client asking hard questions, the useful reading is our guide to what a high score really means, which walks through how to respond when the number is contested rather than how to hide it.

What to do today, in order

Ten minutes of admin beats a week of anxiety.

1. Read the primary source and note the date you read it.
2. Ask your platform whether text leaves your CMS for a model API.
3. Archive your version history somewhere you control.
4. Run the draft through a sentence-level detector and highlight only the flagged lines.
5. Edit those lines for specificity, verify every number and citation against the original, and keep both versions.
6. If disclosure is required, state the tool and the stage plainly.

That is the whole approach: know what the label records, know what it does not, and spend your effort on the sentences a reader will actually judge you by. The ideas, the data and the conclusions have to be yours. No label changes that, and no detector can measure it.

If you want to see which sentences carry the signal before you start editing, the rewrite trial gives you 200 words after email confirmation with no credit card, and detection limits are shown in the current interface.

---

Measurement note: figures in this article come from our own detector run on 2026-09-28 — n = 709 reference passages from guides not marked generated versus n = 12 unedited model passages, median AI score 25% and 34% respectively. Reference authorship was not independently verified; passages are not independent documents. Method: /methodology/.