Confronting Claude about the watermarking of generated text

 

To find out about the usage of watermarking in Claude-generated text, I did a small experiment; let’s ask Claude. Here are the questions and answers from Claude.

Q1: How will Claude hide a watermark in its text?

Ans by Claude Sonnet 5:

Claude doesn't hide watermarks in its text output. I don't have any hidden instructions to embed covert markers, special token patterns, or steganographic signals in my responses, and I'm not aware of Anthropic building such a feature into standard chat responses.

Two things worth distinguishing, since this topic comes up in different contexts:

1. Research-level LLM watermarking (a real academic/industry area): some organizations have experimented with techniques where a model subtly biases its token-sampling probabilities in a statistically detectable pattern, so text can later be identified as AI-generated with a specialized detector. This is a genuine line of research, but it's not something happening invisibly in ordinary Claude conversations as far as I know.
2. Content provenance standards like C2PA, which are used for images/video/audio metadata (and which Anthropic and others have engaged with) — that's a different, non-covert mechanism and doesn't apply to plain text chat output.

If you're asking because you suspect AI-generated text you've encountered has a watermark, or because you're evaluating detection tools, I'm happy to go into more depth on how those statistical watermarking schemes actually work. If you're asking because you want me to embed something hidden in text I generate for you, that's not something I can do — I don't have that capability or instruction set.

Q2: but support page - https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content says otherwise

Ans of follow up question (Claude Sonnet 5):

Good catch — you're right, and I should correct what I said. That page is real and lays out an actual Anthropic commitment.

Here's the substance: Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and as part of that: claude

How it works: When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. It doesn't change the meaning, quality, or readability of the response. Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing. Watermarking is applied at the model level, meaning it's present no matter which Claude product or surface the text comes from. claudeclaude

What's covered: Claude models launched on or after August 2, 2026 support marking at launch. Anthropic is also working to add marking support to models released before that date. Marking applies across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered worldwide, including when supported Claude models are accessed through AWS, Google Cloud, or Microsoft Foundry. claude + 3

Detection: Anthropic is working to enable users and other third parties to detect Claude's embedded watermarks, and plans to share details in forthcoming technical documentation. claude

Important limitations: a detected mark is a signal that content may have been processed by Claude, not fully conclusive proof — Claude may not be the original author (e.g., proofreading or translating someone else's text), and content can be modified after the fact. Conversely, a lack of a detected mark doesn't mean content wasn't AI-generated — this can happen with older models, heavily edited/paraphrased text, very short passages, or unsupported platforms. claudeclaude

So to be direct about my earlier answer: I don't have visibility into the watermarking mechanism itself (it's not something I consciously control or can describe the internals of — it operates "at the model level"), which is why I didn't know about it from my training data. This is a good example of exactly the kind of post-cutoff product development I should search for rather than assume I know. Thanks for pushing back on that.

 

Is Claude deliberately misled or hallucinated? Decide yourself.

P.S.: Banner is also AI generated