OpenAI textGrain Watermarking Explained: What Changes for ChatGPT, Codex and API Users

OpenAI has introduced textGrain, an invisible statistical watermark for AI-generated text. API customers worldwide can opt in starting October 5, 2026, while OpenAI says eligible ChatGPT and Codex text produced for users in the European Union will begin receiving the watermark over the coming weeks. The company is not making text watermarking a global default at launch.

Verified on October 6, 2026 against OpenAI’s announcement and Help Center documentation, the European Commission’s Article 50 transparency guidance, and current Google Search guidance.

What OpenAI announced on October 5

OpenAI’s rollout has three separate parts, and they should not be confused with one another.

The announcement extends OpenAI’s broader provenance system, which already uses Content Credentials and SynthID for supported images and SynthID for supported audio. OpenAI’s public verification tooling remains focused on supported image and audio provenance; text detection is being handled more cautiously because the signal is easier to degrade.

How textGrain works — without hidden characters

textGrain does not add invisible spaces, strange punctuation, hidden Unicode characters or extra watermark-only tokens. Instead, it changes the statistical pattern of the model’s word and token choices while the response is generated.

That distinction matters. Copying and pasting the text does not automatically strip the watermark, because the signal lives in the wording pattern itself rather than in file metadata. But the same property also creates the main weakness: if the wording changes substantially, the detectable signal can weaken.

How reliable is textGrain?

OpenAI is explicit that text watermark detection is probabilistic. It can produce both false positives and false negatives, and reliability depends heavily on passage length, writing domain and later editing.

OpenAI also reports that detection varies by language. In one evaluation across the 24 official EU languages, the company reported 69.0% detection for Spanish and 42.2% for Romanian at a 1% false-positive rate before strengthening the watermark for weaker-performing languages. That variability is one reason a detection result should not be treated like a binary authorship verdict.

What happens after editing, translation or copy-paste?

Copy-paste alone is not the problem. Because textGrain changes token-choice patterns rather than inserting hidden characters, preserving the wording should generally preserve the signal. The bigger issue is transformation.

• Substantial rewriting: can weaken or destroy the detectable pattern.
• Paraphrasing: may reduce reliability because many of the model’s original token choices are replaced.
• Translation: can significantly alter the statistical pattern and reduce detectability.
• Very short answers: may not contain enough signal for confident detection.
• Code and highly constrained text: are harder because there are fewer plausible next-token choices.

OpenAI’s Help Center notes that the EU transparency code does not require watermarks for outputs shorter than 200 tokens — roughly 150 English words — or for code snippets. Coverage also depends on product, model, export path and timing.

Why the EU AI Act is driving the rollout

The European Commission says Article 50 transparency obligations apply from August 2, 2026. Among other requirements, providers of generative AI systems must make AI-generated or manipulated content detectable in a machine-readable way where the law applies. Separate disclosure duties can also apply to certain deepfakes and AI-generated text on matters of public interest.

The EU’s Code of Practice on Transparency of AI-Generated Content is voluntary, but the underlying Article 50 legal obligations are not. The Commission says the code is designed to help providers and deployers demonstrate compliance with those rules.

What changes for ChatGPT and Codex users in the EU?

For most users, the change should be invisible during normal use. OpenAI says textGrain has negligible speed impact and that its benchmark differences with and without watermarking fall within ordinary evaluation noise.

The rollout is also not a new visible badge inside every response. The watermark sits in the statistical pattern of the generated wording. A user can read, copy and work with the output normally.

What changes is provenance: eligible output generated in the EU can carry a signal that an approved detector may later identify. That signal does not expose the user’s account, prompt, organization or conversation, according to OpenAI.

For a broader look at OpenAI’s current developer stack, see our OpenAI DevDay 2026 recap. If you use Codex directly, our guide to GPT-6 Sol, Luna, ChatGPT Work and Codex explains the current model-picker and workflow context.

How API customers can enable text watermarking

OpenAI says API organizations can enable text watermarking either as an organization-level default or as a project-level override. Supported models appear in the relevant settings UI.

1. Open the OpenAI API platform settings.
2. For an organization-wide default, go to Organization settings → Data controls → Text provenance.
3. For a specific project, open Project Settings → Text provenance.
4. Turn on Allow text watermarking.
5. Select the supported model or models you want covered, then save.

Enabling watermarking does not automatically grant access to OpenAI’s text detector. Detector access is a separate program currently limited to approved research and expert organizations.

What a detected watermark does — and does not — prove

textGrain vs SynthID and C2PA

OpenAI now uses different provenance technologies for different media types, because text, images and audio behave differently when they are edited or shared.

OpenAI says it created textGrain to control the trade-off between detectability and response variety. In its internal testing, the method matched or exceeded the text-watermark approaches it compared against, including SynthID for text. The company plans to open-source textGrain, although it has not yet made the public detector generally available.

Does textGrain change Google SEO, publishing or AdSense rules?

There is no evidence in Google’s current Search guidance that an OpenAI textGrain watermark is itself a ranking factor, a spam label or an AdSense approval signal. Google’s documented focus remains on whether content is helpful, reliable, people-first and created primarily for users rather than to manipulate search rankings.

Google’s guidance also says AI or automation disclosures can be useful where readers would reasonably ask how content was created. Separately, using automation primarily to manipulate rankings can violate spam policies. Those are content-quality and intent questions, not a textGrain-specific rule.

For publishers, the safest interpretation is straightforward: provenance technology may improve transparency, but it does not replace editing, fact-checking, source attribution, authorship standards or editorial accountability. A watermark cannot turn low-value automated copy into trustworthy journalism.

Frequently asked questions

Can anyone check whether a paragraph has a textGrain watermark?

Not through a public OpenAI text detector at launch. OpenAI is limiting access to approved researchers and expert organizations while it studies reliability, false positives and false negatives.

Will every ChatGPT answer in Europe be watermarked?

OpenAI says it will add the watermark to eligible ChatGPT and Codex text output in the EU over the coming weeks. Coverage can vary by product, model, output type and length, so “every answer” would be too broad.

Does the watermark contain my prompt or account information?

OpenAI says no. The provenance signal does not encode the user, organization, account, prompt or conversation.

Does copying text into Word, email or a CMS remove the watermark?

Copying alone should not inherently remove the statistical pattern because it is carried by the wording itself. However, extensive rewriting, paraphrasing or translation can make detection less reliable.

Can a watermark prove a student, employee or journalist used ChatGPT?

No. OpenAI specifically warns that a watermark does not identify a person, measure human contribution or establish ownership or responsibility. Detection should not be treated as a standalone disciplinary or authorship verdict.

Bottom line

OpenAI’s textGrain rollout is an important shift from trying to classify AI writing after the fact toward embedding a provenance signal during generation. That is technically meaningful, especially as the EU’s transparency rules move from policy debate into active compliance.

But the limitations matter just as much as the headline. Short passages, constrained writing, edits and translation can weaken detection. The detector is not public. A positive result does not reveal the user or prove authorship, and a negative result does not prove a human wrote the text.

For organizations using the OpenAI API, the immediate decision is whether opt-in watermarking helps meet transparency or governance needs. For ordinary ChatGPT and Codex users in the EU, the biggest change is likely invisible: eligible outputs will begin carrying a machine-readable provenance signal in the wording itself.

You may also like

• OpenAI DevDay 2026 Recap: GPT-6.1 Sol, Agents API and Key Announcements
• GPT-6 Sol vs GPT-6 Luna: Pricing, Benchmarks and Availability
• OpenAI Model Misalignment Framework: What It Means for AI Safety
• ChatGPT Images 2.5: Features, API Pricing and What Changed