Artificial intelligence developer Anthropic has begun introducing machine-readable invisible watermarks into text generated by its Claude models to meet European regulatory mandates.
The new technology embeds an imperceptible signal directly into generated text at the model level, enabling the watermark to survive when output is copied and pasted into other applications.
The implementation responds directly to Article 50 of the European Union Artificial Intelligence Act, which came into effect on August 2, 2026. Under the legislation, providers of generative artificial intelligence systems are required to ensure that synthetic text, image, video, and audio outputs can be reliably identified using machine-readable marking systems.
Systems introduced to the market prior to August 2 have been granted a transitional compliance period extending until December 2. Anthropic confirmed that its new watermarking system will apply globally to all Claude models launched after August 2, rather than being limited strictly to users within the European Union.
The company is also actively working to integrate the text marking technology into its earlier Claude models during the designated transition window.
Anthropic, founded in 2021 by former artificial intelligence researchers, develops the Claude family of large language models. The platform has seen growing adoption in the competitive market for conversational AI tools alongside rivals such as OpenAI's ChatGPT and Google's Gemini.
How model level watermarking functions
Unlike standard visual watermarks or hidden metadata embedded in file headers, the new text marker does not appear as visible disclaimers or concealed text strings within documents. Anthropic explained that the signal is introduced during the text generation process at the model level, causing the marker to travel alongside the words whenever a user copies and pastes a response.
The company has not yet released technical documentation detailing the exact mathematical methods used to encode the signal into Claude's outputs. However, Anthropic stated that it is preparing specialized detection tools to enable third parties and external software systems to verify marked text.
Anthropic noted that the watermarking technique does not provide an absolute or foolproof guarantee of artificial intelligence origin. Significant modifications made to a passage after generation can weaken or erase the embedded signal entirely, while short text snippets may not contain sufficient statistical data for reliable detection.
The absence of a watermark does not conclusively prove that a piece of text was written by a human. Furthermore, users frequently employ Claude to translate, summarize, or edit original human writing. Because those generated responses can also incorporate the model's signal, the mark demonstrates that Claude participated in producing the text rather than proving who wrote the underlying content.
File metadata and open provenance standards
For standalone digital files, Anthropic is taking a different approach because file architectures permit provenance tracking without altering the underlying content. Compatible file formats produced by Claude, including synthetic images, will embed cryptographically signed metadata based on standards from the Coalition for Content Provenance and Authenticity.
Known as C2PA, the open technical framework creates an auditable record identifying which software tool created or edited a file along with its digital modification history. C2PA metadata faces technical constraints because the tracking information can be lost if an image or file is converted to a different format, saved again, or captured through a screenshot.
The C2PA standard was created by a broad industry alliance of media and technology organizations to help combat digital misinformation and establish clear lines of digital content authorship across the internet.
Industry adoption across Google and OpenAI
Anthropic is among several major technology companies deploying content tracking mechanisms to comply with the European Union's new regulatory framework. Google already uses its proprietary SynthID technology to mark text generated by its Gemini assistant by making subtle adjustments to the probability distributions with which the model selects consecutive tokens.
Google's token probability method creates an invisible statistical pattern that readers cannot detect but automated verification software can recognize. For visual content, Google embeds SynthID directly into image pixels, allowing the watermark to remain detectable even after editing, filtering, or compression.
OpenAI currently incorporates C2PA metadata and invisible SynthID watermarks into images created using ChatGPT, Codex, and its application programming interface, while applying SynthID tracking to select audio outputs. OpenAI confirmed plans to expand provenance signals across all modalities, including ChatGPT text outputs, to fulfill its commitments under the European Code of Practice.
OpenAI has not yet disclosed the specific watermarking technology it will deploy or when text marking will launch across ChatGPT outputs.
