On October 5, OpenAI announced it will add invisible watermarks to ChatGPT text generated by ChatGPT and Codexto comply with the EU AI Act’s transparency requirements, which took effect on August 2. This means AI-generated content will be marked at the “native level” for the first time—imperceptible to the human eye but detectable by specialized tools. This is a change all content creators, educators, and developers need to know about.
How is the watermark applied? Not by inserting symbols—but by “reshaping word choice”
This watermark does not insert visible symbols into the text; instead, it uses a technique called textGrain watermarking via logit bias adjustment: a secret key subtly reshapes the model’s word-selection preferences by adjusting the ranking of next-word predictions, embedding a statistical pattern detectable by tools but imperceptible to readers.
Because it resides “within the words themselves,”the watermark persists through copy-and-paste operationsOpenAI states the watermark does not identify individual users, and that model output quality shows “no noticeable degradation” when enabled. This technical report was jointly authored by OpenAI researchers and researchers from the University of Pennsylvania and Yale University.
Launching only in the EU; API developers can use it today
The watermark rollout proceeds in two tiers:
- general usersChatGPT and Codex users in the EU (all plans) ChatGPT will receive the feature gradually over the coming weeks;
- API developerscan manually enable it for select models globally starting today—disabled by default.
Notably, OpenAI explicitly statesWatermarking will not be enabled by default globally for nowThere are commercial concerns behind this—the Wall Street Journal reported in 2024 that OpenAI had already developed text watermarking but delayed its rollout, partly fearing users would switch to competing models without watermarking
Is watermarking truly reliable? It can be removed
OpenAI’s test results are remarkably candid:Watermarks can be disrupted by editingIn one test, replacing 10% of words with synonyms reduced detection accuracy from ~92% to 66%. Additionally, short passages, mathematical answers, and translated text are harder to detect
Due to these limitations, OpenAI has decided to grant detector access only to “approved research institutions and expert organizations” to assess reliability and responsible use cases. The company also explicitly cautions:No watermark ≠ human-written—the text may be too short, heavily edited, or simply generated by another company’s AI
This move comes two months after Anthropic. In August this year, Anthropic announced watermarking for Claude generated text, and implemented itglobally—prompting backlash from some users who argued “I provide the instructions and context; it’s just a tool”Claude OpenAI’s current cautious “EU-only” approach reflects a balancing act between regulatory compliance and user sentiment
What This Means for Ordinary Users
For the vast majority of Chinese users, this change has minimal short-term impact—watermarking currently applies to general users only in the EU. Yet the signal it sends is noteworthy:
- AI content identification is becoming a trend—the EU’s AI Act has taken effect, and regulatory follow-ups elsewhere are only a matter of time
- “AI detection” will become more trustworthy—shifting from probabilistic guessing to key-based verification, with academic integrity and copyright use cases benefiting first
- Content creators should take noteIf you use ChatGPT generated content for publication, it may be identified and labeled in the future.
Progress varies by company: Anthropic globally, Google Gemini toggleable, OpenAI only in the EU
On AI content watermarking, leading companies have taken entirely different approaches:
- Anthropicwas the earliest mover, applying watermarks globally to its models starting in August, Claude Opus 5.5 but this triggered user backlash;
- Google:Gemini’s watermark is toggleable,placing the choice in users’ hands;
- OpenAIis the most cautious, rolling out watermarks to general users only in the EU, with global defaults disabled.
This divergence reflects a trade-off between “compliance obligations” and “user retention.” For content creators relying on AI output, understanding each company’s watermarking strategy helps assess whether published content could be “traced back” as AI-generated. Combined with our site’s curated GPT-6 Sol/Luna evaluation and full-site evaluation list,you can systematically evaluate differences across models in output quality and compliance.
Frequently Asked Questions (FAQ)
ChatGPT What is text watermarking?
It is an invisible tagging technique that subtly adjusts a model’s word-selection patterns to embed a statistical signature detectable by specialized tools—imperceptible to the human eye and不影响 normal readability.
Can watermarks be removed?
Yes, but detection rates drop. Official testing shows replacing roughly 10% of words with synonyms reduces detection accuracy from 92% to 66%. Heavily rewritten or translated text is even harder to identify.
Will this affect Chinese users?
Not in the short term. Watermarking is currently live for general users only in the EU; while API developers can enable it globally, it remains disabled by default. Chinese general users are unaffected for now.
Will watermarking affect ChatGPT ’s response quality?
OpenAI states that model output quality “shows no noticeable change” after watermarking is enabled. It does not identify users and displays no visible markers in the text.
Want to learn more about the inner workings of AI-generated content? Check out our AI Model Library and Tool Comparison Engineor continue reading:Over one-third of web pages are AI-generated · Google Gemini watermark can be disabled · Claude Sonnet 5.5 Review · GPT-6.1 Sol benchmark.