Which AI tools watermark their text?

Two of them do and say so. One built the capability and decided against it. The rest publish nothing at all. Every cell in this table is either a vendor statement, something we found in a sample ourselves, or a report by somebody else, and the page for each one tells you which.

A blank cell means we have no record, not that the assistant is clean. Our samples arrived through a paste path that had already flattened em dashes to hyphens, which means anything invisible was very likely stripped before we ever saw it. Microsoft Copilot's non-breaking hyphens survived only because nothing on that path knew about them. Treat absence here as an absence of evidence, which is exactly what it is.

• No. OpenAI built a text watermarking system and decided not to deploy it.
• Yes. Claude models launched on or after 2 August 2026 carry a SynthID-Text watermark, and no tool can remove it, including ours.
• Yes. Google applies SynthID to text generated by the Gemini app and web experience.
• No published watermark. But its output does carry a character almost nothing else cleans: the non-breaking hyphen.