OpenAI Details New Text Watermarking System for ChatGPT and API
OpenAI announced a new watermarking system for its generative AI tools to comply with the EU AI Act’s transparency requirements.
Key points
- OpenAI is implementing a text watermarking system across ChatGPT, Codex, and its API.
- The move is designed to comply with the EU AI Act, which mandates that AI-generated text be identifiable.
- The system aims to make it easier to distinguish between human-written and machine-generated content.
OpenAI has announced the implementation of a new text watermarking system for its generative AI products, including ChatGPT, Codex, and its application programming interface. The company stated that this initiative is intended to comply with the European Union’s AI Act, which requires providers of generative AI systems to ensure that their outputs are marked in a way that makes them identifiable as machine-generated.
According to a report by 9to5Mac, OpenAI provided details on how it will meet these regulatory obligations. The core requirement of the EU AI Act is that users must be able to recognize when content is produced by artificial intelligence. By integrating a watermarking mechanism into its text outputs, OpenAI aims to provide a technical layer of transparency that supports this legal mandate. This allows downstream users and platforms to detect the origin of the text more reliably than before.
Regulatory Context and Compliance
The EU AI Act represents a significant shift in how generative artificial intelligence is regulated globally. One of its central pillars is transparency, specifically regarding the provenance of AI-generated content. Providers are legally required to implement measures that allow for the identification of AI-generated material. OpenAI’s announcement signals its effort to align its technical infrastructure with these specific legal standards.
The watermarking system is not merely a feature update but a compliance mechanism. It addresses the growing concern over the proliferation of synthetic media and the difficulty in distinguishing it from human-created content. By marking text at the source, OpenAI is attempting to create a verifiable trail that can be used by other systems to flag or categorize AI output. This approach is consistent with the broader industry trend of adopting technical safeguards to meet emerging regulatory frameworks.
Impact on Products and Users
The watermarking system will apply to several key OpenAI products. ChatGPT, the widely used consumer-facing chatbot, will include these markers in its text responses. Codex, the model designed for code generation, will also be subject to the same watermarking protocols. Additionally, the API, which powers many third-party applications and enterprise solutions, will transmit watermarked text.
This broad application means that developers and businesses using OpenAI’s API will need to account for the presence of these watermarks in their own systems. While the specific technical method of the watermarking was not detailed in the initial announcement, the scope covers all major text-generation endpoints. Users of ChatGPT will encounter this system without needing to take any action, as it is integrated directly into the service. For API users, the presence of watermarks may influence how they process, store, or display generated content, particularly if they have their own compliance or verification requirements.
What should you do?
Organizations using generative AI tools should review their content verification policies to ensure they can detect and handle watermarked text. Developers should monitor OpenAI’s official documentation for technical specifications on how the watermarking is implemented in the API. Security teams should assess whether existing content filtering tools can recognize these new markers. Users should remain aware that the presence of a watermark indicates machine origin but does not guarantee the accuracy or safety of the content.
Background: Generative AI
Generative AI creates content from prompts. Agentic AI performs actions using tools. The former outputs text or code; the latter changes system states. Understanding this distinction is critical for security architecture, as agentic systems introduce execution risks that pure generation models do not possess.
Read the full guide: Generative AI vs Agentic AI: Architectural Differences and Security Risks
What to do and how to stay safe: OpenAI
- Review internal content policies to determine how watermarked AI-generated text should be labeled or handled within your organization.
- Monitor vendor documentation for technical details on how to detect or parse watermarks in API responses.
- Verify that your content moderation or filtering tools are updated to recognize new watermarking formats from major AI providers.
- Educate users on the limitations of watermarking, noting that it indicates origin but not truthfulness or safety.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which OpenAI products will have text watermarking?
OpenAI is implementing watermarking for ChatGPT, Codex, and its API.
Why is OpenAI adding watermarks to its text?
The company stated that the watermarking is intended to comply with the EU AI Act, which requires AI-generated text to be identifiable.
Does the watermark guarantee the content is accurate?
No, the watermark serves to identify the content as AI-generated; it does not verify the accuracy or safety of the information.
