As of today, the three biggest chatbots handle invisible text watermarks in three different ways. OpenAI said on October 5 that ChatGPT and Codex will start watermarking generated text for users in the European Union over the coming weeks. Anthropic has been watermarking new Claude models worldwide since August. Google watermarks text in the Gemini app, but its developer story is muddier than you'd expect. If you write with any of these tools, or grade, edit, or publish work that might have been, the details matter more than the headlines suggest.
We read each company's own documentation side by side, and the most useful finding isn't who watermarks. It's that, for now, almost nobody outside these companies can check for a text watermark at all.
Why this is happening now
The trigger is the EU AI Act. Its transparency rules in Article 50 became applicable on August 2, 2026, and they require AI providers serving the EU to mark generated content in a machine-readable way. The European Commission published a Code of Practice on the Transparency of AI-Generated Content on July 31, two days before the rules kicked in, and Anthropic says it was one of about 190 signatories. OpenAI's help center says it signed the same code.
That code has one carve-out worth knowing. According to OpenAI's help center, it doesn't require watermarks in outputs shorter than 200 tokens (roughly 150 words of English) or in code snippets, because there simply isn't enough text for a reliable signal.
Who marks what
All three companies use the same basic idea. When a model picks its next word, it often has several equally good options, and the choice is normally settled by a random number. A watermark swaps that randomness for a secret key, so the text reads the same to you but a detector with the key can spot the pattern. Here's how the implementations differ, based on each company's published docs as of October 5, 2026.
| ChatGPT / Codex (OpenAI) | Claude (Anthropic) | Gemini (Google) | |
|---|---|---|---|
| Where it applies | EU users only, all plans, rolling out over the coming weeks | Worldwide, every Claude surface including the API and Claude Code | Gemini app and web, per Google DeepMind's SynthID page |
| API text | Off by default; any API customer worldwide can opt in per project or org | Applied at the model level, so API output is marked too | Not in official docs; a Google forum staffer said in August that API and AI Studio text is marked, after first saying it wasn't |
| Which models | "Select models" now, legacy models over the coming weeks | Models launched on or after Aug. 2, 2026; older ones being retrofitted | Not specified |
| Method | textGrain, OpenAI's own scheme (it plans to open-source it) | A version of Google DeepMind's SynthID-Text | SynthID-Text |
| Who can check text | Approved researchers and expert organizations, by application | Private-preview detection API for regulators, media, fact-checkers, researchers, educators, EU civil society and obligated enterprises | No public text checker; Google's SynthID Detector covers images, video and audio |
The scope difference is the one most people will notice. Anthropic says it applied watermarking globally because it doesn't yet have a durable way to scope it by region, which drew complaints from some Claude users who felt their own work was being labeled as AI output. OpenAI went the other way and kept the consumer rollout inside the EU, saying the regional approach gives it room to learn from real-world use before deciding anything broader.
How easily the mark breaks
OpenAI published more numbers than its rivals, so we can do a little math with them. In its tests, replacing 10% of the words in a watermarked passage with synonyms dropped the detection rate from about 92% to 66%, and replacing 25% dropped it to 17%. Put another way, a light synonym pass costs more than a quarter of the detections (26 points off 92), and editing one word in four leaves the detector catching fewer than one passage in five.
Language matters too. OpenAI tested all 24 official EU languages at a 1% false-positive rate and found Spanish had the highest detection rate at 69.0% while Romanian had the lowest at 42.2%, a gap of almost 27 percentage points. The company says it can turn up the watermark's strength for weaker languages, and it did so for those below 60%. That same 1% setting is worth translating into real numbers: if someone checked 1,000 essays written entirely by people, they should expect about 10 to come back flagged by mistake. That's a reasonable trade for a researcher studying provenance, and a risky one for a teacher deciding a grade.
Length also cuts both ways. A 150-word email sits right at the code's 200-token floor, while a 1,000-word report is roughly 1,300 tokens, so it gives the detector far more choices to examine. Anthropic adds that factual passages and light proofreading carry very little watermark because the model has fewer free word choices, while a full translation is marked throughout because every word was the model's pick.
What we'd do with this
This part is our opinion. If you're a student or a professional using ChatGPT in the EU, or Claude anywhere, assume long passages you paste in unchanged carry a mark, but also know that none of the three companies says the mark identifies you, your account or your prompt. OpenAI and Anthropic both say outright that a watermark doesn't prove who wrote something, how much a person contributed, or who owns the text.
If you're on the checking side, as a teacher, editor or hiring manager, the honest answer is that you can't run these detectors today, and none of the companies is offering a public text checker. Treat any third-party "AI detector" as a different thing entirely: tools like Pangram guess from writing patterns rather than reading an embedded signal, which both OpenAI and Anthropic point out. And a clean result from any tool doesn't prove a human wrote the piece.
If you build on the OpenAI API and publish generated text to EU audiences, the new opt-in switch under Organization settings, Data controls, Text provenance is worth a conversation with whoever handles your compliance, since OpenAI frames it as a way to meet your own transparency obligations. Developers on Gemini have a harder job, because the only signal that API text is marked so far is a corrected forum reply rather than documentation. We'd ask Google for that in writing before relying on it.
The bigger picture is that text watermarking has arrived as a compliance feature, built for regulators and researchers first. Until detectors open up, it changes less about everyday arguments over who wrote what than the rollout headlines imply.
Primary sources: OpenAI on EU text provenance, OpenAI provenance help article, textGrain technical report, Anthropic on Claude's watermark, Google DeepMind SynthID, and the Gemini API forum thread. Additional reporting: TechCrunch, The Verge, and The Next Web.