Does ChatGPT Watermark Text? OpenAI’s Text Watermark Explained

Last updated: 22 October 2026

OpenAI announced on 5 October 2026 that “over the coming weeks” it would add an invisible watermark to eligible ChatGPT and Codex text in the EU. The mark is a statistical pattern in which words the model picks, so you cannot see it, copy it out or delete it with a “find and replace”. Anthropic’s Claude has watermarked text from newer models worldwide since August 2026, and Google uses SynthID in Gemini.

This guide explains how the ChatGPT watermark works, who it affects, how reliable detection is, what it means for students, businesses and writers, and how other AI tools handle watermarking.

Does ChatGPT watermark its text?

ChatGPT watermarks text for users in the European Union only, as of October 2026. Users in India, the US and other regions receive unwatermarked text in the ChatGPT app, because OpenAI said, “We are not making text watermarking a global default at launch.”

For a broader view of what OpenAI changed in ChatGPT this year, see our roundup of ChatGPT’s new features.

How does the ChatGPT watermark work?

The ChatGPT watermark works by nudging which words the model chooses so that, across a passage, the choices form a statistical pattern only OpenAI’s secret key can detect. Every AI model picks each next word from several likely options; textGrain tilts those choices in a way readers cannot notice.

Because the mark lives in the words themselves, it survives copying and pasting into Word, Google Docs or an email. It does not add zero-width characters, hidden spaces or metadata, so tools that “clean” invisible characters do nothing to it. Developers who want the mark on their own apps can enable it through the API, which our OpenAI API pricing guide covers alongside model costs. According to OpenAI, the watermark does not meaningfully change output quality, with GPT-6 Astra benchmark results broadly similar with textGrain on.

Images work differently. AI image tools typically attach C2PA content credentials, which are signed metadata recording that an image was AI-generated, and Google adds SynthID marks to images. Our guide to ChatGPT image generation covers image labels in more detail.

How accurate is ChatGPT watermark detection?

ChatGPT watermark detection is accurate on longer, unedited text and much weaker on short or edited text, according to OpenAI’s own tests. At a 1 percent false positive target, detection reached about 80 percent on 200-token passages and about 95 percent on 400-token passages.

OpenAI also notes that math answers are harder to watermark because the model has less freedom in word choice, and translation can prevent reliable detection. Most importantly, OpenAI states that a missing watermark does not prove human authorship, and a detected watermark cannot reveal who generated the text, their account, the prompt or how much a person edited it.

Can anyone check if text has a ChatGPT watermark?

No. As of October 2026, OpenAI’s watermark detector is not available to the public, teachers or employers. OpenAI is accepting applications only from approved researchers and expert organizations, granted case by case.

That means the AI detectors you find online cannot read OpenAI’s watermark; they guess from writing style, and they make mistakes, especially on non-native English writing. Read our honest review of the best AI detectors before trusting any score, and never treat a detector result as proof on its own.

What the ChatGPT watermark means for students, businesses and writers

For most people outside the EU, the ChatGPT watermark changes nothing today, but it signals that labelling AI text is becoming normal. The practical impact depends on how you use AI.

Students and teachers

EU students’ ChatGPT text now carries a watermark, but schools cannot check it without detector access. The safer path is to follow your school’s AI policy, use ChatGPT as a tutor rather than a ghostwriter, and disclose AI help when asked. Our study prompts for students focus on learning, and AI for teachers covers fair classroom policies.

Businesses and marketers

If you serve EU customers, assume AI-generated text may be marked and that EU rules increasingly expect disclosure, especially for public-interest content and chatbots. Write a clear internal policy with our AI usage policy template, and keep humans editing and fact-checking every published piece. The ownership question is separate; see can you copyright AI content.

Writers and freelancers

Heavy editing weakens the watermark, but the honest approach is simply to use AI for outlines, research and drafts, then rewrite in your own voice and tell clients how you work. Trying to “beat” watermarks with paraphrasing tools to pass off AI text as fully human can breach client contracts and school rules.

Do Claude and Gemini watermark text too?

Yes. Anthropic’s Claude and Google’s Gemini both watermark AI text, and Claude goes further than ChatGPT by applying it worldwide. Here is how the three compare as of October 2026.

As reported by ThePrint, Anthropic said it applied Claude’s watermark globally because it cannot yet limit it by region, and that it is weaker in code, factual passages and lightly proofread text. Claude-generated images carry signed C2PA credentials. Compare the assistants overall in GPT-6 vs Claude vs Gemini.

Why is OpenAI watermarking ChatGPT text?

OpenAI is watermarking ChatGPT text in the EU to comply with Article 50 of the EU AI Act, which from 2 August 2026 requires providers of generative AI to mark AI-generated text, images, audio and video in a machine-readable format. Fines for breaking these rules can reach EUR 15 million or 3 percent of global annual turnover, whichever is higher.

Article 50 also requires labelling deepfakes and AI-generated text on matters of public interest, and telling people when they are talking to a chatbot. The European Commission published a voluntary Code of Practice in June 2026. These rules are part of a wider push against AI misinformation and scams; our guide to spotting deepfake scams shows why labels matter.

Frequently asked questions

The bottom line

ChatGPT’s watermark is real but limited: EU users get invisible, word-level marks, the detector is not public and editing weakens it. Expect more AI tools to label their output as laws like the EU AI Act take effect, and build habits now: disclose AI help, edit thoroughly and keep your own voice. Policies and rollout details can change, so check OpenAI’s announcements and your school or company rules for the latest position.

Sources

• OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU, BleepingComputer, 5 October 2026, checked 22 October 2026
• OpenAI to watermark ChatGPT text in EU; Claude users in India already getting marked output, ThePrint, October 2026, checked 22 October 2026
• Article 50: Transparency obligations, EU Artificial Intelligence Act, checked 22 October 2026
• ChatGPT release notes, OpenAI Help Center, checked 22 October 2026