ChatGPT Watermark: What textGrain Means for Your Documents

Yes, the ChatGPT watermark exists: as of 24 October 2026, OpenAI is rolling out textGrain to eligible ChatGPT and Codex text in the European Union; it is not a global default for those products, while API customers worldwide can opt in for select models. For a business, the useful response is to document AI use and editorial approval, rather than make a detector score the acceptance test. OpenAI's rollout announcement

Does ChatGPT Have a Watermark? The Product and Region Split

Your product and region determine the answer. ChatGPT is OpenAI's conversational assistant; Codex is its coding agent. The API is how software integrates OpenAI models.

OpenAI announced this on 5 October 2026. Its provenance help page, checked on 24 October 2026, displayed a relative update time within the hour. The announcement remains framed as a rollout, so do not read the table as a guarantee about every response.

For a company with writers in different countries, “we use ChatGPT” is an incomplete process record. Record the product, workspace, generation date and model where available. If you commission work from an agency, ask how it was produced instead of inferring the workflow from your own office location.

What Is ChatGPT Watermark Technology?

textGrain is a statistical pattern in wording. It adjusts the model's selection of tokens, the words or word pieces from which a response is built. A detector with the matching key and settings checks whether the sequence fits that pattern. OpenAI's textGrain technical report

Think of a private rhythm distributed through a paragraph. You can read the paragraph normally, while a compatible checker looks for the rhythm across many choices. That analogy describes a statistical test, not a signature attached to an individual writer.

textGrain inserts no hidden characters, invisible spaces or special punctuation. OpenAI Help Center

ChatGPT Watermark Detection: Who Can Use It?

General business verification is outside OpenAI's current text-detector program. Access is restricted to approved research and expert organizations; enabling API watermarking grants no detector access. OpenAI Help Center

That changes the implementation decision. A procurement manager should not promise an automated textGrain gate for every proposal. An agency should not promise clients an official scan with every handover. First establish that the service is available for your use case and that its terms allow the proposed workflow.

How to Check ChatGPT Watermark Claims

Ask a vendor what its result measures before accepting a “ChatGPT detected” badge. Does it check a compatible embedded signal, or classify writing by its resemblance to AI output? For textGrain, the matching key and configuration matter. OpenAI technical report

OpenAI's public Verify service and Content Provenance API check supported images and audio. Text verification requires separate approval. Content Provenance API documentation

A useful vendor review asks for the supported provenance system, authorized access, input requirements and explanation of uncertain results. If those answers are missing, treat the result as an unverified claim about the document. Do not turn it into a finding against its author.

For an editor, the next question after a suspicious result should be specific: “Can you show the sources and drafting history behind this passage?” That request produces evidence you can evaluate even when no detector is available.

Does Copying From ChatGPT Leave a Watermark?

Preserving the wording can preserve the watermark. Detection is not guaranteed. Light edits may leave a mark intact; substantial rewriting or translation can weaken it. OpenAI Help Center

The operational consequence is that a document's history and its detector result can diverge. A proposal may pass through drafting, legal review, localization and client revision. Its production record should describe those stages even if a later check finds nothing.

Keep editing for accuracy, clarity, confidentiality and the client's voice. Do not define “finished” as “the detector no longer flags it.” That would reward an outcome unrelated to whether the claims are supported or the agreed process was followed.

For example, suppose a consultant uses an approved AI tool to organize interview notes, then writes and checks the final recommendations. The handover should describe that approved assistance accurately. It should not become “entirely human-written” merely because the final document has a different style.

What the Detection Numbers Mean

Published detection results are conditional, not a guarantee for your documents. OpenAI reports approximately 80% detection at 200 tokens and 95% at 400 tokens for content such as psychology, at a target 1% false-positive rate; mathematics performed substantially worse. OpenAI's evaluation

A false positive means a checker flags unmarked text. A false negative means it misses marked text. Neither percentage tells you the probability that a particular employee used ChatGPT. That interpretation also depends on the material submitted, the test conditions and how common marked text is in the population being checked.

For a business, this is a reason to separate screening from a final decision. A short sales email, a technical specification and a long explanatory draft should not all inherit a confidence claim taken from one benchmark. Codex appearing in the rollout also should not become a promise that every code snippet can be reliably checked.

The decision rule is straightforward: if a result could cause rejection, an accusation or a contract dispute, require independent evidence and human review. A numerical score should not displace the work of investigating what happened.

What Article 50 Requires

The EU AI Act separates provider marking from a publisher's disclosure obligations. Article 50(2) requires providers of systems generating synthetic text and other media to make outputs machine-readable and detectable as artificial, subject to technical feasibility and specified exceptions. These include standard assistive editing or changes that do not substantially alter the supplied input or its meaning. European Commission: Article 50

Article 50(4) addresses deployers publishing generated or manipulated text to inform the public on matters of public interest. It requires disclosure, with an exception where there is human review or editorial control and a person or organization holds editorial responsibility. Other exceptions and obligations can also apply. Official Article 50 text

This is not a blanket instruction to put an AI label on every internal memo. Equally, an invisible mark does not settle whether your publication needs a visible notice. Have the person responsible for legal or editorial policy classify the actual use case, including any client terms and other applicable rules.

An editor publishing a public-interest briefing should therefore record the review process and who owns the publication decision. An operations lead circulating an internal planning note has a different question: whether the tool use and data handling meet company policy. Calling both documents “watermarked” does not answer either question.

What Changes for Your Company Documents and Client Deliverables

Change the approval record before changing the writing tool. The useful controls depend on what you promised and who relies on the document.

Owners and operators: name the accountable reviewer

For an internal operating procedure, require the process owner to verify instructions against the way the business works. A polished draft is not evidence that an escalation route, refund rule or service commitment is correct.

Keep a short record of the AI assistance, the sources used and the person who approved the final version. Store it alongside the document under your normal access and retention rules. This gives a future employee somewhere to go when the procedure is challenged.

Editors and agencies: settle the AI-use terms early

For client copy, distinguish permission to brainstorm, permission to draft and permission to edit supplied material. Those are different working arrangements. Agree which are acceptable before production and use those definitions during review.

If a contract promises human-only drafting, escalate any proposed AI use before proceeding. If it permits assistance, record what happened and provide the agreed disclosure. Neither a clean detector result nor a positive result should silently rewrite the contract.

For mixed workflows, ask contributors to identify the stage where AI entered the process. “AI assisted with an outline; the editor verified the claims and approved the final text” is more useful than an unexplained AI percentage, when it accurately describes the work.

Builders and CTOs: record configuration, keep review independent

For an API-based drafting service, assign an owner to the watermarking decision. OpenAI offers model-specific opt-in through organization and project text-provenance settings. OpenAI Help Center

Document the chosen configuration and why it fits the product's obligations. Keep the generation record separate from the release decision: a service may produce marked text while still needing source verification, confidentiality review and an accountable approver.

A small builder can begin with a field in the document's existing approval record. A larger organization can connect that record to its content system. Neither needs to wait for detector access to make AI use reviewable.

What's Overhyped

A watermark is not an authorship verdict. It does not identify an account or reveal prompts, quantify human contribution, establish ownership or verify accuracy. OpenAI also cautions that missing signals do not establish human authorship. OpenAI Help Center

The management mistake is using a provenance question to avoid a quality question. A factually correct, permitted and properly reviewed draft does not become defective because software assisted with it. A false claim does not become acceptable because a detector is silent.

Set standards you can explain without mentioning detection: accurate facts, authorized tools, appropriate handling of confidential information, honest process declarations and accountable approval. Then evaluate any future detector as additional evidence within that process.

Claude and Google's SynthID-Text

Claude's scope differs from OpenAI's EU rollout. Anthropic's current How Claude marks AI-generated content, checked on 24 October 2026, describes worldwide marking from supported models across products, including its API. Text detection is in private preview for eligible organizations, including certain enterprises with verification obligations.

Our Claude watermark guide for business documents covers document-review workflows. For current access, use the private-preview status above; that earlier guide describes the pre-launch position.

Google DeepMind's SynthID-Text also embeds a pattern during generation. Google's SynthID page describes text marking in Gemini's app and web experience. Its public detector is presented for images, video and audio, not as a public textGrain checker.

The implication for procurement is to ask about each provider, model and detector separately. A generic requirement to “check for AI watermarks” leaves the actual acceptance criteria undefined.

A Practical AI-Use Policy and the Monday Move

Make the workflow auditable even when detection is unavailable. Start with one recurring deliverable, such as client proposals or public briefings, and add an AI-use field to its approval record.

1. 1

   Agree the permitted uses

   The owner records whether AI may assist with research organization, drafting, editing or translation. Resolve unclear client requirements before production.

2. 2

   Keep the relevant evidence

   Retain the source material, meaningful draft history and declared assistance under existing confidentiality and retention rules. Avoid copying sensitive prompts into an unrestricted tracking sheet.

3. 3

   Verify and approve the final document

   Assign a person to check claims, sources, required notices and the agreed terms. Record their approval with the final version.

4. 4

   Handle disputes through review

   Ask for evidence, let the contributor explain the workflow, and assess it against the agreed rule. Treat a detector result as an input to that review, never an automatic disciplinary decision.

Suggested policy language, to adapt to your actual agreements:

Record AI assistance when required by the assignment. The named approver remains responsible for checking the final document and meeting disclosure requirements. A detector result alone will not establish human authorship, prohibited AI use or grounds for rejection.

Use the broader AI-generated content watermark policy guide when your workflow also covers images, audio and video. On Monday, the operations owner can pilot the record on the next client deliverable, then ask the editor whether it provides enough evidence to approve the work.

Is there a watermark in ChatGPT?

Yes, within the eligible EU rollout described above. For a particular document, record its production history rather than assuming its status from the product name.

Does ChatGPT watermark text now?

OpenAI has announced and is implementing text watermarking. Use the product-and-region table above when setting a team policy; do not make a universal claim about every response.

How to see ChatGPT watermark in Word?

textGrain is invisible. In Microsoft Word, ask the contributor for the drafting record instead of treating formatting as proof.

How to see ChatGPT watermark in Google Docs?

textGrain is invisible. In Google Docs, keep the declared AI use and approval evidence alongside the document for editorial review.

For practical updates on AI tools and business workflows, join the newsletter.