Depositphotos
According to a 2025 analysis by Ahrefs of approximately 900,000 webpages, nearly 75% of newly published online content now includes some form of AI-generated text. This estimate excludes offline documents and focuses solely on written content. Generative AI is also increasingly being used to create music, artwork, software, photographs, and highly realistic videos. As these technologies advance, distinguishing AI-generated material from human-created work is becoming increasingly difficult, creating significant concerns about transparency and security.
The European Union is addressing these concerns through regulation. Article 50(2) of the EU AI Act requires providers of generative AI systems to ensure that AI-generated or manipulated content can be identified as such.
Anthropic’s Commitment to Transparency and AI Content Watermarking
Anthropic, the developer of Claude, has joined the EU’s Code of Practice on Transparency of AI-Generated Content, alongside roughly 190 other signatories. The company intends to introduce watermarking across Claude’s ecosystem. For text, this will involve an invisible statistical watermark, while supported image formats, including PNG, JPG, and SVG, will contain digitally signed metadata indicating that Claude contributed to their creation.
Depositphotos
Generative AI has expanded rapidly across writing, music, art, software, and nearly every form of digital content. For many users, it has made tasks faster, more affordable, and more efficient. At the same time, its growing use has created serious concerns about misuse and security. These include using AI to produce entire academic essays or violating contracts that specifically prohibit AI-generated content. People may also encounter less serious situations, such as discovering that AI created an impressive song, image, or video without clearly disclosing its use.
The Security Risks of Increasingly Realistic AI-Generated Content
The more significant concern, however, is security. AI-generated content has become increasingly sophisticated, making it harder to distinguish from authentic human-created material. While obviously unrealistic creations, such as a muscular cat moonwalking across the ocean, are easy to recognize as artificial, realistic videos showing public figures promoting violence or fabricated official documents could have serious real-world consequences.
Watermarking and content identification technologies could help address some of these challenges. Anthropic states that its marking system applies to content generated by supported models wherever Claude is available, including the Claude Platform API, Claude applications, Claude Code, Claude Cowork, and Claude Tag. It also covers third-party services that use Claude’s API because Anthropic embeds the watermark directly into the generated output rather than tying it to a specific platform or interface.
Models released from August 2, 2026 onward include content marking from the moment they launch, while Anthropic plans to extend the system to older models in the coming months. This rollout follows a transition period permitted under EU regulations. Although the requirement comes from European legislation, Anthropic is introducing watermarking globally because the company currently lacks a reliable method for restricting the feature to specific regions.
How Anthropic’s Text Watermarking Technology Works
But how does the marking actually work? For text, Anthropic uses a technique based on Google DeepMind’s SynthID-Text technology. Rather than inserting invisible characters or attaching metadata to a file, the watermark is embedded in the way Claude selects words during generation. Large language models typically choose each successive word from several possible options. When multiple words would work equally well in a sentence, the model can select between them using a probabilistic process, and Anthropic’s method uses this selection process to create the watermark.
How SynthID-Based Watermarking Creates an Invisible Statistical Signature
With SynthID-based watermarking, the model’s word choices are influenced by a secret key as well as the text that has already been generated. Over a sufficiently long passage, these subtle decisions form a distinct statistical pattern. The writing remains natural and readable, but it contains an underlying signature that a detector using the correct key can analyze to assess whether the text matches Claude’s watermarking pattern. Anthropic says this approach does not add characters or require additional tokens, does not reduce the quality of the output, and contains no information that could identify a user, organization, or conversation.
For supported file types, Anthropic takes a more traditional approach. When Claude produces files such as PNG, JPG, or SVG images, it adds a cryptographically signed credential to the file’s metadata using the open C2PA standard. The visual content itself remains unchanged. The credential indicates that Claude created or processed the file and does not contain personal information. Compatible software can verify this credential and confirm Claude’s involvement. Since the information is digitally signed, it can also help detect whether the credential or file has been altered after it was created.
Although Anthropic has developed a way to watermark Claude’s text, the company is still working on methods to reliably detect those marks. It plans to introduce a watermark-detection API, but the exact design and functionality have not yet been finalized. For supported files, however, any tool compatible with the C2PA standard can already examine the embedded credentials, and Anthropic says it will also provide its own verification tool.
The Limitations of AI Watermarking and Authorship Detection
Anthropic also emphasizes that users should treat its watermarks as indicators rather than conclusive proof that AI created the content. Detecting a watermark suggests that Claude likely contributed to the material, but it does not prove that a human played no role in its creation. Because Claude embeds the watermark through the words it selects, the system cannot distinguish between content Claude wrote entirely and content Claude helped edit. Anthropic explains that the system cannot determine whether Claude produced the text from scratch or extensively edited human-written material. Likewise, when Claude translates content, the system applies a complete watermark because Claude selects every word in the translated version.
The Challenges of Detecting Watermarks After Extensive Text Modification
Regarding detection, Anthropic acknowledges that minor edits are unlikely to completely eliminate the watermark, but a thorough rewrite that replaces every word can remove it. The company argues that extensive human modification can transform the resulting text to the point where people can no longer consider it largely AI-generated. Extensive rewriting, paraphrasing, translation, or combining the content with other material can also weaken Claude’s statistical watermark until detection systems can no longer identify it. In practice, this means that anyone intentionally seeking to hide Claude’s involvement could potentially do so by significantly rewriting the generated content.
Read more: Google Introduces New Vacation-Planning Tools Across Search, Maps, and Gemini
Read the original article on: New Atlas
