OpenAI's textGrain: The Era of Invisible Watermarks on ChatGPT Outputs

The integration of generative AI models into every area of our lives has made the need to determine the source of content produced in the digital world more critical than ever. Especially in line with the comprehensive AI Act (EU AI Act) put into force by the European Union, tech giants are taking new steps to meet transparency and traceability requirements. Accordingly, OpenAI officially announced a watermarking system called textGrain that embeds digital markers, machine-readable but imperceptible to the human eye, into the outputs produced by ChatGPT and its code-focused model Codex.

This innovation, which will initially cover eligible users within the borders of the European Union, marks the beginning of a new standard for detecting text and code blocks generated by AI.

What Is textGrain Technology and How Does It Work?

Unlike traditional image or video watermarks, marking text-based AI outputs is a very complex engineering problem. Statistical methods are used to give a text an identity without breaking sentence structures or disrupting the reading flow.

Developed by OpenAI, textGrain works by adding an invisible mathematical distribution signal to the model's word and token selection probabilities during text generation. When the model chooses between alternative words that carry equivalent meaning in certain contexts, it creates a hidden statistical pattern according to predetermined rules. This pattern:

• Is completely imperceptible to the human reader and does not affect the literary quality or meaning of the text.
• Can be scanned by specially trained machine learning and analysis tools to verify whether the text was generated with OpenAI's infrastructure.
• Aims to reach accuracy rates similar to or higher than SynthID for Text, the technology Google developed for marking text.

How Editing and Text Length Affect Detection Success

The biggest challenge for text watermarking technologies is preserving the mark when the user copies the output received from the AI and makes changes to it. With the technical data it shared, OpenAI clearly laid out textGrain's limits and performance parameters.

As the chart shows, the system's performance depends directly on two main variables:

1. Text Length (Token Count)

In original texts with no edits (Unedited), the watermark's statistical signal gets stronger as the token count increases. While the detection rate is around 65-70% for short texts of about 200 tokens, this rate rises above 90% at the 400-token level. Since short responses do not produce enough token variety, detection accuracy can drop.

2. Interventions and Word Changes

When the user edits the text, a clear drop in the success rate is observed:

When 10% of the words are replaced with synonyms or rewritten, the detection rate for a 400-token text falls to about 65%.

In scenarios where 25% of the words are changed, the watermark signal almost completely dissipates and the detection rate drops to the 15-20% range.

Because of these limitations, OpenAI stresses that textGrain is not an absolute proof mechanism. The fact that no watermark is found, especially in mathematical formulas, short code snippets or heavily edited texts, does not mean that the text was definitely written by a human. In addition, the watermark system does not certify the accuracy of the text or legal ownership rights.

Scope, Rollout and Privacy Standards

OpenAI is rolling out the textGrain system with a gradual and controlled distribution strategy:

• European Union Users: Starting in the coming weeks, textGrain will be enabled by default for eligible ChatGPT and Codex users on all plans (Free, Plus, Team, Enterprise) in the EU.
• Global API Customers: Developers and enterprise API users worldwide can enable watermarked text generation on selected models on an opt-in basis starting today.
• Cloud Partners: Similar watermark support will be activated in the coming weeks for models accessed through Microsoft Azure and other cloud infrastructures.

The Detection Tool Will Not Be Public

OpenAI will not open the detector tool that verifies the watermarks directly to the general public. Instead, approved academic researchers, regulatory bodies and security experts will be able to access the tool through an individual application and review process.

Clear limits have also been drawn on privacy: the detection tool will only report whether the OpenAI watermark signal is present in the scanned text. It will never reveal which user generated the text, the prompts used or the chat history.

The EU AI Act and Its Effects on the Industry

OpenAI's move aligns directly with the transparency obligations introduced by the European Union's AI Act (EU AI Act). Under the law, major AI developers on the market must mark synthetic content generated by AI, such as text, audio, images and video, in a machine-readable format.

Industry leaders such as OpenAI, Anthropic, Google, Microsoft and Meta are expected to fully comply with these transparency standards by December 2. The textGrain move stands out as one of the most concrete and comprehensive steps taken toward transparently labeling synthetic content across the entire generative AI ecosystem in the future.