Anthropic has announced that it will mark content generated by its Claude chatbot: texts will receive an invisible, statistical watermark, while supported image formats (such as PNG, JPG, SVG) will receive signed metadata indicating AI involvement.
Why the marking is being introduced
As artificial intelligence produces increasingly convincing content, pressure on AI providers has grown to make AI involvement transparent. New European Union rules, which have recently come into force, require providers of generative AI systems to make artificially created or manipulated content identifiable — a requirement that directly relates to Anthropic's announced solution.
How the text watermark works
For text watermarking, Anthropic uses an approach related to Google DeepMind's SynthID-Text method. The technique does not insert hidden characters; instead, it subtly shifts the statistical distribution of words selected during generation. Large language models choose the next word from several plausible options; the watermark slightly tweaks the selection probabilities. A single sentence will not reveal the change, but after thousands or tens of thousands of words the small shifts accumulate into a statistically detectable pattern.
Anthropic stresses that the watermark does not alter meaning or the reader's experience: watermarked and non-watermarked text will be indistinguishable to human readers. However, the watermark can make it possible to detect afterward that Claude likely participated in producing the text. The company also plans to offer a separate watermark-detection API, though it has not shared technical details about its operation.
Anthropic acknowledges the limitations of this approach: the presence of a watermark is an indicator, not definitive proof of AI authorship. Because the watermark is tied to words chosen by Claude, it cannot distinguish authorship from later human editing.
Images: cryptographically signed metadata using C2PA
For images, if Claude produces or processes a supported file (PNG, JPG, SVG), Anthropic will attach a cryptographically signed assertion to the file's metadata according to the open C2PA standard. This is not a visible “made by AI” label on the image itself; it records in the metadata that Claude created or processed the file. The signed assertion does not include personal data, and the cryptographic signature can help reveal whether the assertion or the file was later tampered with.
Rollout schedule and impact
Anthropic says models released on or after August 2 will support marking at launch. The company is working to add the feature to older models during the EU's permitted transitional period over the coming months. Although the requirement is European, Anthropic will apply watermarking globally from the start, citing a lack of a workable region-by-region restriction.
According to Anthropic, watermarking will have a negligible effect on model latency, will not generate extra tokens, and will not increase the cost to serve or use the models.
Summary
Anthropic's step addresses regulatory and transparency pressures by applying statistical watermarks to text and cryptographically signed metadata to images, and by planning a detection API. The measures can indicate Claude's likely involvement but do not constitute conclusive proof of sole AI authorship, nor do they affect the experience of human readers.
