Claude and ChatGPT now watermark their text: what changed and what it means

Two of the biggest AI companies now put an invisible watermark in the text their chatbots write. Anthropic started with Claude models released from August 2, 2026, everywhere in the world. OpenAI announced on October 5, 2026 that ChatGPT and Codex text in the EU will be watermarked over the following weeks. Neither watermark is a hidden character you can find and delete. Both are patterns in word choice.

What Anthropic announced

On August 11, 2026, Anthropic said that all Claude models released on or after August 2, 2026 watermark the text they write. The mark is applied in the model, so it is present whether the text comes from the Claude apps, the API, Claude Code, Claude Cowork or Claude Tag. It applies worldwide and cannot be switched off, and Anthropic said it would extend the mark to older models as well (TechCrunch).

Images and other files Claude creates get signed provenance metadata using the C2PA Content Credentials standard. Unlike the text mark, file metadata does not survive a screenshot or a plain-text copy.

What OpenAI announced

On October 5, 2026, OpenAI said it will add an invisible watermark to text from ChatGPT and Codex for users in the EU, on every plan, rolling out over the following weeks. Developers anywhere can turn the same watermark on in the API for select models. It is off by default there. Detector access starts with approved researchers and expert organizations only (OpenAI, TechCrunch).

OpenAI built a text watermark years ago and held it back. In 2024 it pointed to how easily rewriting could get around it and the risk of stigmatizing non-native English speakers who use AI as a writing aid. The EU rules changed the calculation, at least for Europe.

Why now: the EU AI Act

Both moves trace back to Article 50 of the EU AI Act, whose transparency rules took effect on August 2, 2026. It requires providers of AI systems that generate text, images, audio or video to mark the output in a machine-readable way, so it can be detected as AI-made. Anthropic signed the EU code of practice on marking AI-generated content in July 2026, alongside other major AI companies. Anthropic chose to apply its mark worldwide; OpenAI is starting with the EU.

How a statistical watermark works

A language model writes by picking one token at a time from a list of likely options, and often several options are about equally good. A watermarking model uses a secret key to tilt those close calls in a pattern. One choice tells you nothing, but over hundreds of words the tilt adds up to a signal that a detector holding the key can measure from the text alone.

Because the signal is spread across the words themselves, it survives copying, pasting, retyping and changing the file format. It weakens as the words change. OpenAI says that in one test, replacing 10% of words with synonyms dropped detection from about 92% to 66%, and that short passages, math answers and translated text are harder to detect. Anthropic says its mark may persist through some editing.

Hidden characters are a different thing

Before 2026, "ChatGPT watermark" usually meant something else. In April 2025, researchers noticed that longer answers from OpenAI's o3 and o4-mini models contained narrow no-break spaces (U+202F) where normal spaces belonged. OpenAI said they were a quirk of training, not a watermark. Characters like these, along with zero-width spaces, citation markers and tracking tags on links, really do ride along in copied AI text, and they are easy to find and remove.

That is the job of a text cleaner like our free AI watermark remover. It strips hidden characters and formatting. It cannot touch a statistical watermark, and a tool that claims to remove one in a click is removing characters, not the watermark.

Who can detect these watermarks?

As of October 5, 2026, Anthropic has said it plans verification tools but has not released a public text detector. OpenAI is limiting detector access to approved researchers and expert organizations while it studies reliability. Google, which has marked Gemini text with SynthID since 2024, offers detection for content from its own models. Expect this to change: the EU rules push providers toward detection that others can use.

Mainstream AI detectors such as GPTZero and Turnitin work differently. They judge writing style rather than reading a watermark, and that method has its own error rates. How AI detectors work covers the details.

What it means if you use AI for school

A watermark is evidence about where words came from, not a verdict on how they were used. People use Claude and ChatGPT to edit and translate their own writing, and both companies say short passages are hard to judge. Still, if your school bans AI on an assignment, assume that Claude text written since August 2026 can be identified once detectors are available.

The practical takeaways: check your school's policy, keep your drafts and version history, and if you use AI for help, make the final wording your own. Writing you have genuinely rewritten carries far less of the original model's pattern, and it is also the version you can explain and defend.

Where Write It fits

Our AI watermark remover is free and runs in your browser. It removes hidden characters, citation leftovers, tracking tags and markdown, and says plainly what it cannot remove. The Humanizer rewrites AI-flavored wording. We do not promise that any tool defeats a watermark or a detector, because nobody outside the AI companies can test against their keys.