Claude Adds Invisible Text Watermarks Under EU AI Act (Aug 2026)
Future Claude models embed an undetectable watermark in generated text to comply with the EU AI Act, effective August 2, 2026. The watermark doesn't change output quality, speed, cost, or user privacy—but it makes Claude's involvement in a document statistically detectable.
Future Claude models now embed an invisible watermark in the text they generate. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, along with several other major AI providers and around 190 total signatories. The EU AI Act requires AI providers to mark AI-generated content, with the requirement taking effect August 2, 2026. Anthropic is applying watermarking globally at launch because it doesn't yet have a reliable way to limit it by region.
How it works: same words, different randomness
Claude generates text one word at a time. When multiple words would work equally well—say, overcast versus grey to finish "The weather today was cold and…"—Claude picks one using randomness. Normally, that randomness comes from an arbitrary number generator.
With watermarking, the randomness comes from a secret key instead. The words Claude picks still feel natural and random to any reader. But someone with the key can examine the sequence of word choices and calculate the probability that Claude generated the text. Claude uses a version of Google DeepMind's SynthID-Text approach, published in a Nature paper in 2024, which itself builds on a 2022 proposal by Scott Aaronson. Nothing is added to the text—no hidden characters, no extra tokens. Only the source of randomness changes.
What changes for you: almost nothing
In internal testing, Anthropic found no impact on content, creativity, or readability. Google DeepMind ran a live test on a portion of Gemini traffic and found no statistically significant difference in thumbs-up/thumbs-down ratings between watermarked and unwatermarked responses. Watermarking adds no extra tokens, so Claude costs the same to run.
The watermark also carries no identifying information. It cannot be traced to a specific user, organization, or chat session. Anthropic is building a watermark detection API—details are still being finalized—but that API will only tell you whether Claude was likely involved in producing the text, not who asked.
When the watermark signal is weak
Watermarks need many free word choices to build a detectable pattern. Three situations reduce the signal:
Short text. Fewer choices means less evidence. Confidence grows with passage length.
Factual or constrained text. When only one word is correct—"Isaac Newton's most famous work was called Principia Mathematica"—the watermark has nothing to act on. Code, where exact syntax is required, is similarly sparse.
Human text Claude lightly edited. The watermark only lives in words Claude chose. If Claude fixed five typos in a 1,000-word document, those five changes may not be enough to detect Claude's involvement.
Detection is always a probability, never a certainty.
Images use a different system
When Claude produces an image file (.png, .jpg, or .svg), it attaches a C2PA content credential—a small, cryptographically signed note in the file's metadata stating that Claude was involved. C2PA is an open industry standard also used by camera manufacturers and photo-editing software. Nothing in the image itself changes; the credential sits in the metadata. Like the text watermark, it includes no identifying information about the user.
What you now know
Future Claude models embed an invisible text watermark to comply with the EU AI Act (effective August 2, 2026).
The watermark changes the source of randomness in low-stakes word choices—not the words themselves. Quality, speed, and cost are unchanged.
No user or organization is identifiable from the watermark. It only signals Claude's probable involvement.
The signal is weakest on short text, factual passages, code, and lightly edited human writing.
Images use a separate C2PA metadata credential, not an embedded watermark.
A detection API is coming; results are probabilistic, not proof.