ChatGPT adds hidden text signatures in Europe, OpenAI complies with new rules

How resistant is the TLDR

watermark to modifications?

Internal tests showed that after replacing 25% of words with synonyms, the test success rate only dropped to 17%.

Why are Detection Tools Private?

Access rights are limited to approved scientific institutions and expert organizations.

Get 3 free stock e-books

OpenAI fully deploys textGrain invisible statistical watermark in Europe

OpenAI is deploying an invisible statistical watermark called "textGrain" in ChatGPT and Codex output in Europe. The move aims to meet transparency compliance requirements in the EU AI Act.

Internal testing showed that when 25% of words were replaced with synonyms, the test success rate only dropped to 17%. Currently, access to Detection Tools is still limited to approved scientific research institutions and expert institutions.

Although this is mandatory across the EU, API users around the world can voluntarily enable the watermark by setting it.

Technical details and implementation background

On October 5, OpenAI announced plans to embed invisible watermarks in ChatGPT and Codex responses in EU territory. The company said the move was a direct response to transparency obligations under the EU Artificial Intelligence Act.

We are expanding our content traceability approach to cover the text domain to respond to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. Our tools can already help verify that an image or audio file was created by us... -- OpenAI (@OpenAI) October 5, 2026

The technology behind the initiative is called "textGrain." It creates a hidden statistical signature by subtly influencing the word selection process of the language model. The accompanying detection application then analyzes the text to identify the signature. According to OpenAI, this approach does not rely on hidden characters, invisible spaces or unusual punctuation marks.

As a result, typical circumvention strategies, such as removing hidden characters from copied text, cannot eliminate the watermark. Signatures are woven directly into the language pattern itself.

How resistant is the watermark to modification?

The effectiveness of the test varies depending on the length of the content. OpenAI uses "tokens" to quantify text-fragments of words interpreted by language models. About one token is equivalent to three-quarters of standard English words.

• In an internal evaluation, the detection system successfully identified approximately 66% of the unmodified responses, which were approximately 150 words long.
• When the content reaches about 300 words, the success rate increases to about 92%.

Modifying the content quickly reduces the effectiveness of the watermark. For a 300-word paragraph, replacing 1 synonym in every 10 words reduces the detection accuracy from 92% to 66%. When 1 out of every 4 words was replaced, the detection rate plummeted to just 17%. OpenAI's internal findings suggest that once text is edited, compressed or rewritten, the system faces significant challenges.

The company also pointed out that the implementation of watermarking has little impact on the performance indicators of its latest model, GPT-6 Astra.

Why are Detection Tools Private?

OpenAI chose not to make the detector widely available immediately. Instead, access will be granted only to vetted researchers and specialized institutions.

The company clarified that positive test results will not disclose user identities, input prompts or conversation details. Negative results also lack certainty. Short, modified or translated content may evade detectors without triggering alarms.

Content generated by other AI platforms will also not activate the system because it specifically searches OpenAI's proprietary signatures. OpenAI made it clear that failure to detect a watermark does not confirm the identity of a human author.

Visual and audio content follow separate verification processes. Users can now upload these file types to OpenAI's public verification platform to scan their SynthID watermarks.

For regions outside the EU, ChatGPT watermark is disabled by default for ordinary users. However, API clients around the world can activate this feature for specific models through their project or organizational preferences.

This deployment follows the transparency provisions of the European Union's Artificial Intelligence Act that have been implemented since August. OpenAI positions the watermark as part of its ongoing response to regulatory requirements.

The company also cited broader industry data in its statement. Research by TRM Labs shows that criminal exploitation of AI tools has increased by 40% year-on-year.

OpenAI has not disclosed when or whether Detection Tools will be made available to the public. Currently, the scope of implementation in the EU and limited access to researchers represent the existing scope of the initiative.