WatermarkRemoverPro

ChatGPT vs Claude: how the two text watermarks differ

Two of the largest model providers now mark their text, but not in the same way or in the same places. The differences decide what a reader, an employer or a checking tool can actually conclude.

Scope: regional versus worldwide

OpenAI’s watermark covers ChatGPT and Codex in the EU, rolling out over the coming weeks, with an opt-in for API developers anywhere. Anthropic announced its Claude watermark in August 2026 and said it applies worldwide.

Both sit under the same driver, the EU AI Act’s transparency rules, which took effect on 2 August 2026. Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have committed to the EU’s code of practice on AI-generated content, so more providers are likely to follow.

Method and robustness

OpenAI’s textGrain is a keyed bias on next-word choice and comes with a published report and robustness figures: a 10% synonym replacement dropped detection from about 92% to 66%. Both vendors’ marks belong to the family of word-choice watermarks, which weaken under editing and translation.

OpenAI reported no meaningful change in model performance with the mark on. Neither mark is visible to a reader, and neither identifies the user.

Detection and what it means for a checker

Neither provider publishes a detection key. OpenAI is limiting detector access to approved researchers and expert organisations; a third-party checker therefore cannot test either mark. WatermarkRemoverPro tests the keys it holds, including a public reference key for self-testing, and lists them on every result.

Mixed provenance is the usual case: a draft from one model, edited by a person, summarised by another. Because a mark shows only that a system touched part of a passage, the honest reading of any single result is narrow. See /guide/claude-ai-watermark and /guide/chatgpt-watermark for each mark on its own.

Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. WatermarkRemoverPro's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.

Answers, in full

Questions people actually ask

Which is harder to remove, the ChatGPT or Claude watermark?
There is no published head-to-head comparison, and we cannot measure either because neither key is public. OpenAI’s own figure for its mark is a drop from about 92% to 66% detection when 10% of words are replaced with synonyms.
Do other AI companies watermark text too?
Google, Meta and Microsoft are among the companies that have committed to the EU’s code of practice on AI-generated content, alongside Anthropic and OpenAI. Check each provider’s own announcements for what they have actually shipped.