OpenAI Outlines Invisible Watermarks for AI-Generated Text

OpenAI follows Anthropics footsteps and has described a method for identifying text generated by ChatGPT through an invisible watermark. The approach is part of the company’s work on text provenance, or the ability to determine where content came from.

The watermark works by making a small, deliberate adjustment to the way a language model selects words. The resulting writing is intended to remain natural to readers, but the pattern of word choices contains a statistical signal that can be detected with a separate system. The signal is not a visible label or an added phrase, so a reader would not normally notice it in the output. This approach is conceptionally similar to Anthropics method of watermarking text and paves the way for standardizing AI governance

Claude Introduces Invisible Watermarks and Signed Metadata for Enhanced Transparency

Claude, the AI model developed by Anthropic, has recently implemented a significant advancement with the introduction of invisible watermarks embedded in all text outputs and signed metadata on files. This evolution not only enhances transparency but also aligns with global regulatory standards, especially the EU AI Act’s Code of

CakadoCan Karsten

According to OpenAI, the method is designed to preserve the quality of generated text while supporting detection at scale. A detector would look for the relevant pattern across a sufficient amount of text rather than relying on a single word or sentence. This distinction matters because the watermark is a property of the overall generation process, not a mark that appears in a fixed location.

OpenAI also describes limitations. Text provenance methods can be affected when generated writing is substantially changed. Rewriting, paraphrasing, translation, or other forms of editing may weaken or remove the signal, depending on how the text is altered. Short passages can also provide less material for reliable statistical detection. As a result, a watermark detector would not establish the origin of every passage with certainty, particularly when the content has been modified.

The company presents watermarking as one possible part of a broader provenance system rather than a complete answer on its own. The method concerns the text produced by a model and does not by itself explain how that text was used, edited, or combined with human writing. Those distinctions limit what a detection result can show. Nevertheless its an important step towards AI governance and safe use

Sources:

Our approach to EU text provenance rules

How OpenAI is approaching text watermarking under EU rules. Learn where watermarks apply, how detection works, and why access starts with researchers.

OpenAI