What is AI text watermarking, and what does it mean for you?

How a text watermark works

A language model writes one token at a time, and often several next words are about equally good. Normally a random number settles the choice. A generation-time watermark changes where that randomness comes from: a secret key plus the preceding few words decide the choice instead. The text still reads normally, but it now follows a pattern that someone with the key can test for.

Google DeepMind's SynthID-Text, described in Nature in 2024 and released as open-source code in Hugging Face Transformers, is the best-documented method. Google's developer documentation describes detection as probabilistic, with three possible outcomes: watermarked, not watermarked, or uncertain. Anthropic says the watermark on Claude's text is a version of SynthID-Text.

Why it is showing up now

Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic audio, image, video or text to mark outputs in a machine-readable format that is detectable as AI-generated, as far as technically feasible. It exempts assistive functions for standard editing and outputs that do not substantially alter the input. The European Commission says the marking obligations applied from August 2, 2026, and that about 190 organisations had signed its Code of Practice on transparency of AI-generated content by the end of July 2026, naming Anthropic, Google, Meta, Microsoft, Mistral and OpenAI among signatories of the provider section.

Signing the code does not tell you how each company marks text. Anthropic has published details; for other providers, check their own documentation rather than assuming they use the same method.

What Anthropic says about Claude

Anthropic's announcement of August 14, 2026 and its help center article, both read on September 26, 2026, make these claims. They are the vendor's statements, not independent measurements.

• Nothing is added to the text: no hidden characters and no extra tokens, so price and speed are unaffected.
• The watermark carries no information about the user, organisation or chat.
• It applies at the model level, across the Claude apps, Claude Code, the API and cloud partners such as AWS, Google Cloud and Microsoft Foundry, worldwide.
• The help center table lists text watermarks for Fable 5.1, Mythos 5.1, Opus 5.5 and Opus 5, with older models to be covered by December 2, 2026.
• Code carries less watermark, because exact tokens leave few free choices; comments can carry it.
• Files such as PNG or JPEG images get a signed C2PA Content Credential in their metadata, which is a separate mechanism from the text watermark.
• Detection is in private preview for eligible organisations such as regulators, researchers, media and educators, and for enterprises with compliance obligations.

What a watermark cannot tell you

A watermark check answers one narrow question: how likely is it that this specific model, with this specific key, chose these words? Anthropic lists the limits plainly. It cannot confirm that a human wrote the text, cannot detect other AI models, works poorly on short passages and light edits such as proofreading, and cannot tell Claude wrote this apart from Claude heavily edited this. A complete rewrite removes it.

That makes watermark detection different from AI detector services, which guess from writing style without a key. Neither is a reliable basis for accusing someone. Our guide on whether AI detectors work covers why style-based detection produces false positives, and a missing watermark should be read the same way: as absence of one signal, not proof of human authorship.

This answer does not cover how to remove or evade watermarks. Stripping provenance marks is outside what RepoRadar recommends.

What builders should check

Anthropic states that it saw no impact on quality in internal testing, and the SynthID-Text paper reported no significant quality difference in Gemini user ratings. A separate question is whether the same sampling change alters what an agent does. In a study published September 17, 2026, security firm Lasso applied SynthID-Text's non-distortionary configuration to open models such as Llama-3.1-8B, Gemma 3, phi-4 and Qwen3-4B. On the Berkeley Function Calling Leaderboard, an average of 6.5 percent of tool-call verdicts changed between watermarked and unwatermarked runs, even when net accuracy barely moved. Under a fixed prompt-injection technique, refusal on harmful requests weakened on some models, and the effect varied with the watermark key.

Those results come from a vendor with a security product to sell, on open models with keys the researchers chose, not on Claude or any provider's production key. Treat them as a reason to test, not as proof that your agent is affected.

• Re-run your agent evaluations and red-team prompts when a provider enables watermarking on the model you use, or changes models.
• Compare runs item by item, not only headline accuracy, because changes in opposite directions can cancel out.
• Keep argument validation and approval gates on consequential tool calls, whatever the model.
• If you ship AI-generated content into the EU, check what Article 50 requires of you as a deployer; Anthropic's help center says builders should assess this independently.

Limits of this answer

RepoRadar did not test watermark detection or measure any model's behavior. Claims about Claude come from Anthropic; the agent-behavior findings come from one vendor study that has not been independently replicated as far as we know. This is not legal advice about the EU AI Act. Provider coverage and detection access are changing, so check the dated sources below.

A useful next action

If you only use AI to write, nothing changes day to day: disclose AI use where your school, employer or publisher requires it, since the watermark does not do that for you. If you build on a hosted model, check its provider's marking documentation, then schedule a paired evaluation of your agent with watermarking on, using the steps in our answer on evaluating an AI agent before you trust it in production.