OpenAI Launches textGrain Watermarking for EU Compliance

OpenAI will introduce an invisible “textGrain” watermark to qualifying ChatGPT and Codex texts within the European Union over the coming weeks. (textGrain: an invisible watermark that adds a statistical signal to the word choices of a language model.) This rollout aims to satisfy transparency requirements under Article 50 of the EU AI Act, which mandates marking and detection for AI systems by December 2, 2026. If you are currently operating without a system to identify AI-generated content in the EU, this implementation provides a path to meeting that December 2 deadline.

EU AI compliance officers monitoring the December 2 deadline

EU-based AI compliance officers should monitor the December 2, 2026 deadline for marking and detection obligations under the EU AI Act. OpenAI’s textGrain deployment provides a mechanism to meet these transparency rules, but the system’s effectiveness is limited by the length of the text and the nature of the content. For instance, the system shows much lower detection rates for math than for psychology, and it does not require watermarking for free-form text shorter than 200 tokens. (tokens: the basic units of text, such as words or parts of words, processed by language models.)

Compliance officers should update their internal AI usage policies to reflect the presence of watermarks on EU-generated content. They can also compare the detection rate of OpenAI’s textGrain against other methods, such as Google Gemini watermarking, to assess the reliability of their provenance checks. (provenance: the history and origin of a piece of content to verify its source.) Because the system is only currently available in the EU, compliance teams must ensure their internal rules account for this regionality.

Legal counsel for content agencies drafting authorship contracts

Legal counsel for content agencies should revise service level agreements to remove reliance on watermarks as a definitive proof of human authorship. OpenAI stated that a missing watermark does not prove a person wrote the text, and the watermarking does not measure how much a person contributed to the final output. Instead, legal teams should rely on human-in-the-loop workflows to verify content authenticity. (human-in-the-loop: a workflow involving human intervention or oversight to verify the authenticity of automated content.)

To manage risk, legal counsel should monitor the false positive rate of the textGrain system during the initial EU rollout. (false positive rate: the frequency at which a system incorrectly identifies non-AI content as being generated by artificial intelligence.) Because the target false positive rate is 1%, large volumes of text can still produce a significant number of errors at scale. Consequently, the presence or absence of a watermark cannot serve as a reliable legal audit for individual contributions in a contract.

AI researchers studying textGrain detection rates

AI researchers can submit a research access request to OpenAI to use the text detector for provenance testing on a case-by-case basis. The detector is currently restricted to approved researchers and expert organizations because OpenAI noted that the risk of missed watermarks and false positives makes it unsuitable for the general public at launch. Researchers can use the ELI5 dataset to conduct independent tests on the effectiveness of the watermark.

Researchers should monitor the textGrain detection rate for 400-token passages, which OpenAI reports to be approximately 92%. They can also compare the performance of textGrain against SynthID for text, as OpenAI claims the textGrain method matched or exceeded the performance of that specific method during testing. (SynthID: a specific method for watermarking text used for identifying the origin of content.)

Growth engineers integrating OpenAI API models

Growth engineers integrating the OpenAI API should note that watermarking is off by default and requires a manual opt-in for select models. This allows developers to choose when to enable provenance for their end-users while maintaining the ability to scale without the watermark. Since the rollout is currently limited to the EU, engineers must architect their systems to handle region-specific watermarking logic.

Detection reliability and the synonym bypass risk

OpenAI’s textGrain method works by adding a statistical signal to a model’s word choices rather than a persistent static identifier. The system targets a 1% false positive rate and has achieved about 80% detection in 200-token passages and about 95% in 400-token passages for content like psychology. However, the reliability of the watermark drops significantly when the text is edited.

Editing the text weakens the signal by changing the probabilistic word choices. OpenAI reported that replacing 10% of words with synonyms reduced detection from approximately 92% to 66%. If the replacement increases to 25% of the words in 400-token English passages, the detection rate drops further to 17%.

Comparison of detection rates

Baseline (no editing)

Approximately 92% detection for 400-token passages

10% Synonym replacement

Approximately 66% detection

25% Synonym replacement

Approximately 17% detection

Regionality and the Anthropic comparison

OpenAI is using the European market as a sandbox to refine the watermark before potential global expansion. This contrasts with Anthropic, which marks text from supported Claude models worldwide because it does not yet have a durable way to scope watermarking by region. This means a company using both providers may face inconsistent provenance capabilities based on geography.

Who this is not for: General consumers in the US or Asia

General consumers in the US or Asia can ignore this rollout as it is currently limited to the European Union; these users should follow general OpenAI product updates for other features. Creative writers seeking absolute authorship proof should skip this and instead focus on establishing manual human-in-the-loop workflows, as OpenAI stated that watermarks cannot measure individual contribution.

Owen’s scenario: A content agency with dual offices

Owen runs a content agency with offices in Berlin and Toronto that both use the same OpenAI plan. Because the Berlin office is in the EU, the 400-token articles they generate will be watermarked via textGrain. However, the Toronto office will generate the same articles without a watermark because the rollout is currently restricted to the EU.

Owen cannot use a single global policy to determine the “AI-generated” status of his content based on detection results. If Owen switched his entire workflow to Anthropic’s Claude, both offices would receive marked text because Anthropic does not scope watermarking by region.

What this can be used for

What this replaces: OpenAI's textGrain replaces manual AI detection heuristics

OpenAI’s textGrain replaces manual AI detection heuristics by providing a standardized statistical signal for identification. This allows for more consistent detection compared to human-designed rules that may not account for the probabilistic nature of word choices in large language models.

Reported by Search Engine Journal on 6 October 2026: OpenAI To Watermark ChatGPT Text In The EU, Opens API Opt-In via @sejournal, @MattGSouthern.