OpenAI's new text watermark can add useful provenance evidence, but it cannot decide who wrote a passage, whether the passage is accurate, or whether a person contributed meaningful work. The practical takeaway is simple: treat a watermark result as one narrow signal inside a documented review, not as a verdict.
On October 5, 2026, OpenAI published its approach to the European Union's text provenance requirements. Eligible ChatGPT and Codex outputs in the EU are scheduled to receive an invisible text watermark, while API customers can opt in globally for supported models. The rollout matters because teams will increasingly encounter claims that a watermark either proves AI authorship or proves human authorship. Neither conclusion is supported.
What OpenAI introduced
OpenAI calls its text watermarking system textGrain. It changes statistical patterns in word selection while aiming to preserve meaning and quality. It is not a set of hidden characters, invisible spaces, or a label that a reader can see.
The initial system has important scope limits:
- The ChatGPT and Codex rollout applies to eligible text generated in the EU.
- API customers can enable watermarking for supported models.
- Text verification is initially restricted to approved researchers and expert organizations.
- OpenAI's public verification tool does not provide general text verification at launch.
- Short passages and code are harder to evaluate, and some short or code-heavy outputs may be exempt.
Those limits mean many organizations will not be able to submit ordinary text and receive an authoritative OpenAI result. A sound process must therefore work even when direct verification is unavailable. The current Help Center also excludes general business verification and academic-integrity checks from the text-detector program. This framework does not imply that an editorial team, employer, or school is eligible for access.
OpenAI text watermark verification has three valid outcomes
The safest interpretation is a three-state decision, not a binary AI-versus-human judgment.
| Review result | What it supports | What it does not support | Next action |
|---|---|---|---|
| Signal detected | The text likely passed through a supported watermarked generation path | Accuracy, ownership, responsibility, prompt identity, or the absence of human contribution | Preserve the result and review the surrounding workflow evidence |
| No signal detected | No supported signal was found in the tested material | Human authorship or proof that AI was not used | Check eligibility, length, editing, translation, age, and provider coverage |
| Not testable or inconclusive | The available method cannot support a determination | Any positive or negative authorship claim | Record the limitation and rely on process records, disclosure, and source evidence |
A detector can also report a false positive or miss a watermark that is present. A result therefore remains evidence to interpret, not a definitive classification.
This distinction matters because watermark detection can weaken after editing, paraphrasing, or translation. OpenAI also reports lower performance on some types of text. In its own controlled tests, detection varied by passage length and domain. These figures describe an evaluation setting, not a universal benchmark for real-world documents.
A four-step provenance review
Use this process before acting on any text watermark claim.
1. Preserve the original record
Save the submitted text, its source location, timestamps, file history, and any relevant conversation or revision records. Do not begin by pasting the only copy into multiple tools. Remove sensitive data before sharing material with AI. Preserve what was actually received.
2. Check whether the text was eligible for a signal
Ask four narrow questions:
- Was the text generated after the relevant rollout?
- Was it produced through a supported OpenAI product, model, account setting, and region?
- Is the passage long enough and prose-like enough for the method to be meaningful?
- Was it edited, translated, summarized, or combined with other writing?
If these facts are unknown, record them as unknown. Do not convert missing context into certainty.
3. Interpret the result narrowly
Write the finding in evidence language:
A supported provenance signal was detected in this version of the text.
or:
No supported provenance signal was detected. This result does not establish human authorship.
Avoid labels such as "AI-written," "cheating," "fraudulent," or "human-written" unless separate, verified evidence supports that conclusion and an authorized reviewer is responsible for making it.
4. Corroborate with workflow evidence
Compare the signal with evidence that answers different questions:
- Draft history can show how the work developed.
- Source notes and claim verification can show where factual claims came from.
- Prompt and revision records can show how AI was used.
- Human review records can show who checked accuracy and approved publication.
- A disclosure can explain the role of AI without pretending to quantify authorship.
Watermark evidence becomes more useful when it is paired with a clear process record.
Worked example: reviewing a 350-word briefing
This is a hypothetical interpretation exercise, not a claim of detector access or a tested NEXAIUM result. Imagine a reviewer examines a 350-word market briefing together with a result from an authorized research evaluation that reports a watermark signal.
The wrong conclusion is: "OpenAI wrote this document."
The supported conclusion is: "The submitted version contains a signal associated with a supported OpenAI text-generation path."
The team then checks the writer's source notes, revision history, and disclosure. The records show that an AI system produced an initial outline, while the writer selected sources, rewrote the analysis, checked every claim, and approved the final copy. The watermark result does not measure those human contributions. It also does not confirm that the market facts are correct.
The review can now separate three questions:
- Provenance: Was a supported AI generation path involved?
- Process: What did the person create, revise, verify, and approve?
- Accuracy: Are the claims supported by reliable sources?
One signal cannot answer all three.
When a negative result proves very little
An absent signal may have several explanations. The content may predate the rollout, come from a different provider, fall outside the supported scope, be too short, contain code, or have been substantially edited or translated. OpenAI also notes that watermark detection is probabilistic and can miss watermarked text.
This makes negative-result language especially important. "No signal detected" is an observation. "Written by a human" is a separate claim that the observation does not establish.
A practical policy for teams
Organizations adopting provenance checks can use five rules:
- Never make disciplinary, legal, hiring, or publication decisions from a watermark result alone.
- Record the tested version, method, date, reviewer, and known scope limits.
- Keep provenance, authorship, contribution, ownership, and accuracy as separate questions.
- Give the person involved a route to provide drafts, sources, and process evidence.
- Recheck provider documentation before relying on product availability or verification access.
The last rule is essential because rollout scope, supported models, regional coverage, and verifier access can change.
Who should use this framework
This review is useful for editorial teams, educators designing fair processes, compliance teams, product leaders, researchers, and organizations building AI-use policies. It is not a substitute for legal advice, a provider-specific verification procedure, or a documented investigation process.
The larger lesson is durable: provenance technology can strengthen an evidence trail, but it should not collapse a complex human process into a binary label.
Source notes
- OpenAI, Our approach to EU text provenance rules, published October 5, 2026.
- OpenAI Help Center, Provenance signals in OpenAI-generated content, verified October 12, 2026.
Facts and product scope were verified on October 12, 2026. Availability, supported models, regional coverage, and verification access may change. Check the current provider documentation before implementation.