Do AI chatbots watermark their text?

The persistent folk theory is that AI watermarks are hidden characters — zero-width spaces or special Unicode tucked between the words. That is not how any deployed text watermark works.

A real text watermark is statistical. As the model writes, it faces many small choices where several words would do equally well. A watermarking scheme nudges those choices toward words on a secret list that changes with the surrounding context. Any single choice is unremarkable; across a few hundred words, the pattern of choices becomes measurable to someone who holds the key, while remaining invisible to everyone else. The text contains only ordinary characters, reads normally, and looks identical whether or not the mark is present.

This is why the watermark cannot be seen, and also why it cannot be stripped the way formatting can. It is not in between the words. It is the words.

Invisible Unicode characters are a real phenomenon — they routinely survive a copy-paste out of a chat window, and the invisible characters guide covers where they come from. But they are a side effect of web rendering, not a deliberate mark. You can list and remove every one of them with the Invisible Character Scanner, and the statistical watermark, if one is present, will not have been touched.

The EU AI Act's transparency rules (Article 50) began applying on August 2, 2026. Providers of generative AI systems must mark AI-generated content in a machine-readable way where technically feasible, and the major labs signed a Code of Practice describing how. The vendor picture as of August 2026:

Google Gemini — yes, live. Google's SynthID has watermarked Gemini's text output since well before the deadline, alongside its image, audio, and video marks. Google signed the transparency code in July 2026 and is pushing SynthID as a cross-industry standard, with partners including Apple, NVIDIA, ElevenLabs, and OpenAI.

Anthropic Claude — yes, for new models. Anthropic signed the same code. Claude models launched on or after August 2, 2026 embed a text watermark from day one, applied at the model level rather than per-product. Older Claude models are in a transition period. Anthropic says detection tooling for third parties is coming but has not shipped publicly yet.

OpenAI ChatGPT — no, not for text. As of August 2026, ChatGPT does not watermark its text. OpenAI has deployed marks in other modalities — SynthID in generated images since May 2026, and marks in voice output, plus C2PA provenance metadata — and has researched text watermarking for years without shipping it.

Two things follow from that picture. First, "this text has no watermark" does not mean "no AI wrote this" — the most-used chatbot in the world does not mark its text at all, and older models never will retroactively. Second, the situation is moving quickly; this section is dated because it will age.

Alongside text watermarks, generated files — images, PDFs, SVGs — increasingly carry C2PA provenance metadata: a signed label saying which tool produced the file. Unlike statistical watermarks, that metadata is fragile by design and disappears on any re-save, screenshot, or format conversion.

Because the watermark is woven into word choice, it behaves differently from anything formatting-based:

Copy-paste: survives. The mark travels with the text into any editor, form, or document. Pasting as plain text, "removing formatting", or laundering the text through Notepad does nothing to it — those operations strip styling and hidden characters, and the watermark is neither.

Light editing: mostly survives. Fixing typos, swapping a handful of words, or deleting a paragraph leaves most of the marked choices intact. The signal degrades gradually with each change rather than disappearing at some threshold.

Heavy rewriting, paraphrase, translation: degrades badly. Providers say this themselves — the published limitations note that heavily edited, paraphrased, or translated text may no longer carry a detectable mark. By the time a human has genuinely rewritten a passage, most of the original word choices — and therefore most of the mark — are gone.

Short text: too little signal. A sentence or two does not contain enough choices to measure. Watermark detection is a statistical claim and needs length to reach confidence.

Detection: gated by the provider. Reading the mark requires the provider's detector. Google offers SynthID detection; Anthropic's is announced but not yet public. The commercial "AI detectors" schools and editors use do not read these watermarks at all — they guess from style, which is a far weaker signal with a well-documented false-positive problem, discussed in the AI tells guide.

One more limitation the providers state plainly: a detected mark means the text passed through a model, not that the model authored the ideas. Ask Claude to proofread your own essay and the output can carry the mark. Provenance is not authorship, in either direction.

If you paste AI output into documents: nothing about your workflow changes. The watermark does not affect how text renders, breaks no software, and adds no characters. The things that do break documents — smart quotes in code, stray em dashes, zero-width characters, Markdown symbols in Word — are ordinary formatting artefacts, and cleaning them up with the AI Text Cleaner is unrelated to watermarking.

If you are wondering whether a text you received is marked: you cannot currently check, unless it came from Gemini and you use Google's detector. There is no public tool that reads Claude's mark yet, and ChatGPT text has no mark to read. Anyone selling a universal "AI watermark detector" for text in 2026 is selling something else.

If you are tempted by "watermark remover" tools: be skeptical twice over. Against statistical watermarks, a character-level cleanup provably does nothing, and the vendors who claim otherwise cannot test their own claim — the detectors are not public. The only operation that reliably degrades the mark is a deep rewrite, which is just a paraphrasing model wearing a different label, with the quality loss that implies — and output rewritten by a marking model simply carries that model's mark instead. Our tools remove formatting artefacts and report what they find; they do not and cannot remove watermarks, and we would rather say that plainly than sell you a myth.

If your concern is being flagged: the watermark is the least of it, because reading it requires the provider's cooperation. What actually gets text flagged today is the visible layer — the mechanical and stylistic tells that style-based detectors and human readers key on. That layer is real, checkable, and covered honestly in the seven signs guide.