Anthropic's Claude Fable 5.1, released September 1, 2026, now embeds an invisible statistical watermark in all generated text worldwide. The mark can estimate whether Claude was involved in a passage, but it cannot identify you, your account, or your conversation. It is the first Claude model launched after the European Union's August 2, 2026 AI Act transparency cutoff, and the watermark applies everywhere the model runs, not only in Europe.

What Can the Watermark Actually Prove?

The watermark works by changing how Claude makes word choices. When multiple words are equally good options, the model normally picks one at random. Under watermarking, a secret key plus the preceding few words determines the choice instead. Anyone holding the key can test a passage and assign a probability that Claude was involved.

However, the watermark has significant limitations. It cannot answer several critical questions that people often assume it can:

• User Identity: The watermark carries no user, organization, or chat information, so it cannot identify who generated the text or which account produced it.
• Conversation Source: One key covers all Claude output globally; there is no per-session variant, so the mark cannot trace text to a specific conversation.
• Human vs. AI Authorship: A negative result proves nothing; other AI providers use different keys or methods, so absence of Claude's mark does not prove a human wrote the text.
• Proofreading vs. Generation: The mark lives only in words Claude chose; a few corrections are too little to register, so it mostly cannot detect text that Claude only proofread.
• Code Quality: Exact tokens leave no free choices, so the watermark is weak on source code, though comments and documentation can carry it.

Why Is This Regulation, Not a Product Choice?

The watermark is not an Anthropic innovation; it is a compliance requirement. Article 50 of the EU AI Act requires providers to mark AI-generated output in a machine-readable way. In July 2026, Anthropic signed the accompanying Code of Practice alongside roughly 190 other signatories. Non-compliance carries fines of up to EUR 15 million or 3 percent of global turnover.

Anthropic applies the mark worldwide because, it says, it has no durable way to scope it by region yet. The watermark uses a version of SynthID-Text, a method Google DeepMind published in Nature in 2024. Nothing is added to the text and there are no hidden characters, so copying and retyping preserve the mark at no extra token cost.

How to Understand Watermark Detection and Removal?

• Who Can Check: Only Anthropic and organizations in a private-preview detection API can test for the watermark. These include regulators, law enforcement, media, fact-checkers, researchers, educators, civil society groups, and enterprises with their own Article 50 obligations. No consumer-facing checker exists today.
• Removal Method: The only way to remove the watermark is to rewrite every word. Rewrite every sentence by hand or with another model, and nothing is left for the key to read. However, rewriting through another cloud model swaps one provider's provenance for another's, since Google, Meta, Microsoft, Mistral, and OpenAI all signed the same Code of Practice.
• Local Generation: The only rewrite with no key involved is one on hardware you control. Models like Kimi K3, Qwen, DeepSeek, and Gemma have downloadable weights, allowing generation under no key to begin with.

Detection confidence rises with passage length. The EU Code of Practice does not require marking free-form text under 200 tokens, roughly 150 words, so the watermark is unreliable on short text. Files work differently; supported image types get a signed C2PA content credential in their metadata, readable by any C2PA-aware tool and stripped by converting, re-saving, or screenshotting the file.

What Does This Mean for Claude Users?

Claude Fable 5.1 and Claude Mythos 5.1 are one model with two safeguard settings. Anthropic kept input and output pricing where Fable 5 left it, but cut cache-read pricing by 75 percent. Earlier Claude models launched before the August 2 cutoff, including Opus 5 (July 24), Sonnet 5, and Fable 5 (June 9), will receive the watermark over the coming months.

A positive watermark result means Claude processed the text, not that Claude authored it. Where disclosure matters, users should be transparent about AI involvement. Where control matters, generating text on local hardware avoids any watermark entirely. The watermark carries no privacy risk because it contains no identity information, only a statistical pattern that indicates AI processing occurred.