OpenAI Adds Watermarks to ChatGPT Texts: Mandatory in the EU, Optional in the API
· AI · Cem Koyluoglu
OpenAI is adding invisible watermarks to ChatGPT texts to comply with EU rules. Unlike Anthropic, the company will allow API customers worldwide to disable this feature. According to the reports...
TL;DR OpenAI is adding an invisible watermark to ChatGPT and Codex outputs to satisfy the EU AI Act, offering it as an optional feature for API users and through partners like Microsoft Azure. The watermark, called textGrain, is designed to be detectable in most text—especially longer passages—but can be bypassed with editing or short math content. Tests show it does not noticeably affect output quality.
Key Highlights • OpenAI introduces textGrain, an invisible watermark to meet EU AI Act labeling requirements. • The watermark is optional for API customers and will be available via Microsoft Azure. • Detection is strongest in longer passages and weaker in math or heavily edited text. • No noticeable impact on model performance, and the detection tool is initially limited to selected researchers.
OpenAI Adds Invisible Watermarks to ChatGPT Texts to Meet EU AI Act Requirements
OpenAI announced that it will embed invisible watermarks in the text generated by ChatGPT and Codex to comply with the European Union’s Artificial Intelligence Act. Unlike Anthropic, which has made its watermark mandatory for all users of Claude, OpenAI will allow API customers worldwide to opt out of the feature. The watermark will also be offered through cloud partners such as Microsoft Azure in the coming weeks.
How the Watermark Works
The EU AI Act requires AI providers to label machine‑generated content in a machine‑readable format. OpenAI’s solution, called textGrain, inserts an invisible statistical signal into the model’s word choices. The approach is similar to Google’s open‑source SynthID watermark used by Claude. OpenAI plans to release textGrain as open source, enabling others to build upon it.
Detection Accuracy and Influencing Factors
OpenAI reports that detection rates depend heavily on text length and subject matter. With a 1 % false‑positive threshold, the watermark was detected in approximately 95 % of a 400‑token passage about psychology. When the passage was shortened to 200 tokens, detection fell to about 80 %. Detection is noticeably lower in mathematical content, where the model’s word‑choice freedom is more constrained. Longer passages strengthen the statistical signal and improve detection, though OpenAI has not provided data to confirm this hypothesis.
Impact of Editing on Detection
Editing the text can reduce detection rates. Replacing only 10 % of the words with synonyms lowered detection from roughly 92 % to 66 % in a 400‑token passage. Altering a quarter of the words dropped detection to about 17 %, making it easier to bypass the watermark. This vulnerability could lead users who wish to avoid watermark detection to switch to models that do not offer the feature.
Effect on Output Quality
OpenAI conducted tests on its flagship model, Astra, and on eight evaluation benchmarks, including GPQA Diamond, BrowseComp, and DeepSWE. The tests showed no noticeable difference in performance with the watermark enabled or disabled. However, the studies did not conclusively determine whether the watermark affects writing quality, a point that critics have raised regarding Claude’s watermark as well.
What the Watermark Reveals
OpenAI emphasizes that a detected watermark does not disclose any information about human creativity, editing, ownership, responsibility, or the correctness of the text. It merely indicates that the content was produced by an AI system. The absence of a watermark does not prove that a human wrote the text, as short, edited, translated, or otherwise altered content may evade detection.
Limited Access to the Detection Tool
Initially, the textGrain detection tool will be available only to selected researchers and expert organizations that apply through a form. OpenAI will evaluate access on a case‑by‑case basis under the EU Code of Practice. The tool will simply report whether a watermark is present; it will not reveal user identities, prompts, or conversations. OpenAI plans to broaden access once it is confident that the results can be interpreted responsibly, though no timeline has been announced.
Other Public Verification Tools
Existing verification services, such as the openai.com/verify page and the Content Provenance API, will remain publicly accessible.