Which AI Tools Actually Watermark Your Text and Code (and What That Means for You)

Anthropic just started watermarking everything Claude generates. That raised an obvious question: is every AI tool doing this now, or just Claude? Here’s the actual state of things, and what it means if you use any of these tools day to day.

What Claude’s watermark actually is

Every Claude model launched after August 2, 2026 embeds an invisible mark in what it generates, whether that’s regular text or code. It’s not visible on the page and it doesn’t change what Claude wrote, it’s a statistical pattern in the specific word and token choices Claude made along the way. Think of it like a passport stamp: it shows your text or code passed through Claude, it doesn’t prove Claude wrote every word of it from scratch. If you draft something yourself and just have Claude tighten it up, the mark still shows up.

The trigger was the EU AI Act, specifically the transparency rule (Article 50) that became enforceable August 2, 2026 for newly launched AI systems. Anthropic rolled the change out worldwide, not just in Europe, across the Claude Platform API, claude.ai, Claude Code, Claude Cowork, and Claude accessed through AWS, Google Cloud, and Microsoft Foundry.

Does editing remove it?

Light edits, a word swap here or there, won’t touch it. The pattern is spread across hundreds of choices in the piece, not concentrated in a few words. A genuine rewrite, changing the structure and the wording (including having a different AI rewrite it for you), does erase the signal. Very short pieces don’t carry enough information for reliable detection either way.

One thing worth knowing: Anthropic hasn’t released the actual detection tool yet, only said more technical detail is coming. So right now, nobody outside Anthropic can check whether a given piece of text carries the mark.

What about ChatGPT, Gemini, and the rest?

OpenAI has had a working text watermark for ChatGPT for years, reportedly around 99.9% effective in testing, and has chosen not to ship it. The stated reasons: a meaningful share of surveyed users said they’d use ChatGPT less if it shipped, real concerns about paraphrasing defeating it anyway, false positives on real human writing, and a specific worry about unfairly flagging non-native English speakers who use it as a writing aid.

Google already watermarks AI-generated images and audio through a system called SynthID, and it’s become close to an industry default. OpenAI itself adopted SynthID for images and audio in 2026 rather than building its own.

Meta has a similar invisible watermarking tool for AI-generated video.

xAI’s Grok stamps a visible “GROK” watermark on images it generates (whether you see it depends on your subscription tier). Whether it does anything comparable for text is genuinely unclear, the available reporting on that specific point is thin, so take it as an open question rather than a confirmed fact.

Open-weight models (Llama, Stable Diffusion, and similar) are a different case entirely: because anyone can download and run the model themselves, no company can guarantee a watermark stays in place, since it can just be stripped out of the code before running it.

What this actually means for you

You might be thinking about switching tools over this. You don’t need to. Google, Meta, and OpenAI have all separately committed to the EU AI Act’s Code of Practice, so more of this is coming across the board regardless, Claude just moved first on text specifically.

If you use AI for real work, text or code, the practical takeaway is simple: a light polish pass doesn’t hide that a tool touched your work, at least not on Claude right now. What the mark actually proves is narrower than people assume too. It shows a tool was involved, not how much of the final piece is yours versus the model’s. And since nobody outside Anthropic can currently check any of this, none of it is enforceable against you today, it’s worth understanding for later, not panicking about right now.