OpenAI is set to implement machine-readable text watermarking in the European Union to comply with the EU AI Act. Over the coming weeks, invisible watermarks will start appearing in some responses from ChatGPT and Codex.

We’re expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model. In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Customers using our API can turn on text watermarking for select models worldwide today.

— OpenAI (@OpenAI) October 5, 2026

However, the company acknowledges that the technology currently lacks the capability to reliably determine the origin of any given text.

Text Watermarking Details

In the coming weeks, an invisible watermark will be added to suitable text content from ChatGPT and Codex for users in the EU. This marking will be available across all subscription plans.

The technology, named textGrain, incorporates a statistical signal into the text based on the model's word choice. A specialized detector then checks for the presence of this signal in the text.

OpenAI clarifies that this initiative is driven by the EU AI Act, which mandates generative AI providers to make generated text machine-readable for identification purposes. The company emphasizes that current text watermarking methods are still in their infancy.

Challenges in Identifying AI-Generated Text

According to OpenAI, the effectiveness of textGrain is influenced by the length and content of the text. With a false positive rate of 1%, the system was able to detect the watermark in about 80% of texts containing 200 tokens and approximately 95% of texts with 400 tokens during tests involving psychological questions.

In contrast, the detection rates for mathematical texts were considerably lower due to the reduced variability in word choice.

Another challenge arises from editing; in tests with 400-token texts, replacing 10% of the words with synonyms decreased watermark detection from about 92% to 66%, and replacing 25% of the words dropped the detection rate to 17%.

Source: OpenAI.

As a result, OpenAI is not currently making the detector openly available. The company has begun accepting applications from researchers and expert organizations to test the technology and assist in evaluating its reliability.

Limitations of the Watermark

OpenAI has warned that the presence of a watermark does not automatically prove authorship by AI.

The signal does not indicate the extent of human contribution, the legality of its use, or who is responsible for it. It also does not identify the user nor disclose their queries or conversation history.

Conversely, the absence of a watermark does not imply that the text was written by a human. The signal can be lost due to editing, translation, or if the text is too short. Additionally, the material may have been created by a different model or prior to the implementation of watermarking.

Opt-In Watermarking for API Users

Simultaneously, OpenAI is launching optional text watermarking for API users globally. Starting today, clients can enable watermarks for specific models, although this feature will remain off by default.

The company also plans to release the technology as open-source, allowing other developers to incorporate it into their projects. In the coming weeks, OpenAI aims to extend watermark support to models available through cloud partners.

However, the company does not intend to make textGrain a universal standard for ChatGPT at launch. Initially, OpenAI wants to test the system in the EU and gather data on its performance in real-world scenarios.

Expanding Content Provenance Systems

The text watermarking initiative will be part of a broader content provenance system that OpenAI is developing to trace the origins of materials.

For images, the company already employs Content Credentials and C2PA standards, along with invisible watermarks known as SynthID.

OpenAI believes that relying on a single method is insufficient: metadata can be stripped away, and text can be relatively easily rewritten, translated, or edited. Therefore, the company intends to combine various technologies and gradually refine them.

As a reminder, Spotify recently introduced an AI Persona label for AI performers starting in September.

Follow ForkLog on social media

Telegram (main channel) Facebook X