OpenAI Makes ChatGPT textGrain Watermarks Default in EU
OpenAI will enable textGrain watermarks by default for ChatGPT in the EU, but tests show editing 10 percent sharply weakens detection reliability in trials.
Summary
OpenAI announced on October 6, 2026, that ChatGPT will automatically watermark generated text in the European Union within weeks. The proprietary system, textGrain, responds to the EU AI Act, effective since August, which requires machine detectable markings on AI content. In ChatGPT and Codex, watermarking will remain optional and off by default outside the EU. The API will also offer it without enabling it automatically. Anthropic, by contrast, enabled text watermarking globally in August.
textGrain embeds patterns in word choices that are imperceptible to readers and are not intended to meaningfully reduce output quality. A specialized detector with a key can identify them, but OpenAI will initially restrict access to selected researchers and organizations, adding others through an approval process. OpenAI published a technical paper, but its tests show only 92 percent detection in the best case. Changing 10 percent of an output reduces successful detection by almost 30 percent, while changing 20 percent lowers it by almost 75 percent. Shorter and translated text also performs worse than longer text. Existing approaches including SynthID and C2PA are similarly vulnerable to circumvention by users with basic technical knowledge.
Positives
- EU users will receive automatic textGrain watermarking in ChatGPT within weeks, supporting compliance with the EU AI Act.
- 92 percent best case detection gives textGrain a measurable ability to identify unedited ChatGPT output.
- textGrain embeds machine detectable patterns without visibly altering wording or meaningfully reducing general output quality.
- Researchers and organizations can gain detector access initially, with an approval process planned for additional users.
- Users outside the EU can optionally activate watermarking even though it remains off by default.
Risks & concerns
- Changing 20 percent of an output lowers textGrain detection success by almost 75 percent in OpenAI's tests.
- Changing only 10 percent of generated text reduces successful detection by almost 30 percent.
- Short and translated outputs have lower detection rates than longer text, limiting reliability across common uses.
- 92 percent best case detection means textGrain cannot reliably identify every watermarked output.
- SynthID, C2PA and similar watermarking systems can be circumvented by users with basic technical knowledge.
- Detector access will initially be restricted to a limited number of researchers and organizations.