Access the blog post here.

The text and media generated by Claude will now have: 

• A watermark for text responses: “Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.” 
• Provenance metadata for media, such as .svg, .png, or .jpg, following C2PA open standard: “This metadata follows the Coalition for Content Provenance and Authenticity (C2PA) open standard, which is used across the industry to record information about content provenance. If a signed metadata label is present, it signals that a file was processed by Claude and lets you detect whether the file has been tampered with.”

Watermarking and provenance are implemented worldwide: “Marking will apply to output from supported models wherever Claude is offered, worldwide,” clarified Anthropic.

It also said that “Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch.” 

Detection tools and technical documentation to be published soon: Anthropic has yet to enable users or other third parties to detect Claude’s embedded watermarks and provenance metadata. It also said that it will release detailed technical documentation soon. 

said on that

“People’s first concern shouldn’t be watermarking, it should be AI detectors!” said Alex Cui, an Machine Learning reseacher and founder of GPTZero, a AI-text detection company. He said it in the context of circumventing the watermarking mechanism. He further laid out four challenges for the frontier AI labs to overcome: 

1. Their watermark needs to work token-by-token because they are streaming their text to users. Many watermark methods plan sentences or paragraphs at a time, or change the text after its entirely written, in order to make their watermark robust to paraphrasers, and a frontier lab cannot afford to do this yet (SemStamp, PostMark)

2. If the secret key leaks, the watermark is busted. To avoid a large blast damage from this, you need to have a couple secret keys in rotation.
3. There are some texts, like code, that cannot be arbitrarily changed, otherwise the code will break. In those cases, the watermark needs to selectively change words in parts of the text that can tolerate synonyms (i.e. like variable naming) – see SWEET, EWD, Invisible Entropy.
4. They will need to educate their users on how to deal with false positives and false negatives of a detector, which is a big challenge (one we put a lot of effort into)

Anthropic implemented these measures to comply with the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. The EU finalised the Code of Practice containing these labelling guidelines in June 2026.

Article 50 of the EU AI Act has two sections aimed at two different stakeholders:

• Providers: Those who build AI systems; and
• Deployers: Those who use these systems to create and publish content.

In its “Code of Practice on Transparency of AI-Generated Content,” the EU laid out the symbols that AI platforms can use to label AI-generated content. They are: 

• “AI GENERATED” — for fully AI-generated deepfakes or published text.
• “AI MODIFIED” — for partially AI-manipulated content.
• Basic “AI” icon — a minimal version that can be supplemented with a custom interactive layer.

Read MediaNama’s lowdown on the EU’s Code of Practice for AI-generated content here. 

MediaNama’s Take:  

• If platforms comply with EU rules but not India’s, does this suggest weaker enforcement by Indian authorities? Or does it simply reflect the global adoption of EU standards? The Ministry of Electronics and Information Technology released India’s synthetically generated information (SGI) rules in February 2026, while the European Union finalized its code of practice for generative AI content in June 2026. Most platforms and intermediaries operating in India have not fully complied with India’s SGI rules, except for users’ self-disclosures. 

During a recent Parliament session, Member of Parliament Karthi Chidambaram asked for figures on complaints, takedowns, or requests under the new three-hour takedown notice for SGI content, but the government did not disclose them.

• Platforms need stronger anti-circumvention measures for watermarking and C2PA metadata: Platforms need to strengthen C2PA metadata and watermarking to prevent bypassing. For example, developers created tools to circumvent Google’s SynthID after its launch. How can AI companies and intermediaries design C2PA standards that resist circumvention, especially when metadata is stripped via screenshots or shared on encrypted platforms like WhatsApp or Signal? How much metadata survives when AI-generated content is modified or doctored?

“As of now, there is not a single AI model that can detect a really good deepfake-generated video and tell you this is fake,” said Tarun Wig, co-founder of Innefu Labs, at the India AI Impact Summit (2026). 

• Can platforms technically trace the originator of AI-generated media, especially when law enforcement agencies request or legally compel them to do so? What are the law enforcement compliance requirements for C2PA, and how would platforms provide user information for the original prompt-giver if requested? 

“Every AI developer requires a sign-in, and the data the user inputs is embedded in the metadata. Through that metadata, you can trace back the person who created the content.” — a speaker at MediaNama’s discussion on ‘Regulating Deepfakes in India.’

• Watermarking has limited impact on deepfake content: Major AI platforms already use watermarking or C2PA standards, yet deepfakes persist. Recently, the Press Information Bureau’s fact-checking account debunked a deepfake video of Prime Minister Narendra Modi, Finance Minister Nirmala Sitharaman, businessman Mukesh Ambani and Rajya Sabha MP Sudha Murthy, where Sitharaman’s deepfake promoted a fraudulent scheme promising users 1 lakh rupees. Despite the video being AI-generated, the fact-checking post on X did not include an “AI-generated” label.

“The biggest problem with deepfakes is that [something new is] generated every time, and the content can change quite a bit, as opposed to the classical photographs that we have. These techniques are based on the idea that there’s a database of hashes that we can cross-check with really quickly. And so it’s really hard for a company to maintain a database of hashes of every possible content that can be generated,” — Gautham Koorma, machine learning researcher at MediaNama’s Deepfakes and Democracy. 

• Dilution of AI-generated content labels: If most online content is/will be AI-generated, whether in text or media formats, a universal “AI-generated” label risks losing its purpose. 

Also Read: 

NOTE: The article was updated on August 12 at 9:56 am to include GPTZero founder’s comments.