Claude to Incorporate Invisible Watermark in AI-Generated Texts

Anthropic, the company developing the artificial intelligence assistant Claude, will implement an invisible watermark in texts generated by its models, according to La Derecha Diario. The measure, which responds to the European Union’s AI Act, will be applied globally and will be permanent and irreversible, according to an official company statement cited by the outlet. This decision has generated concern among users who use the platform to draft content, especially in academic and professional settings where AI use is restricted.

What the watermark is and how it works

The watermark is an imperceptible statistical pattern inserted into the sequence of tokens generated by the model. It was developed by Google DeepMind and published in the journal Nature in 2024, under the name SynthID-Text, according to La Derecha Diario. Anthropic adopted this technology to comply with Article 50(2) of the European AI Act, which requires identifying AI-generated content. The implementation alters the randomness in word selection but does not use hidden characters or store personal information from conversations.

The watermark survives copying, pasting, or exporting the text to other platforms, including the Claude API, Claude Code, Claude Cowork, and Claude Tag, as well as in cloud services such as AWS, Google Cloud, or Microsoft Foundry. It will also affect visual files such as PNG, JPG, or SVG, which will incorporate cryptographic metadata under the C2PA standard.

What changes it implies for users

The measure is already active for new models and will be integrated into pre-existing tools after December, with a transition period for models released before August 2, 2026, according to La Derecha Diario. This means that any text generated by Claude will carry a digital signature that can be detected using a verification tool that Anthropic will launch soon. Detection will be binary and high-probability, allowing to know whether the model participated in the writing.

The watermark will drastically affect translations but presents vulnerabilities in very short texts, programming code, or strictly factual content. Additionally, if the user only corrects minor punctuation or grammar errors, the model’s intervention might not leave a trace. However, a comprehensive rewrite of the content could eliminate the signal, as the company itself admits.

Reactions and practical consequences

Given this scenario, many users have decided to abandon the platform and migrate to alternatives such as ChatGPT or Grok, which do not implement watermarks, according to La Derecha Diario. The warning is especially strong for professionals, content creators, and students who use Claude for work with zero-AI policies. Concerns focus on privacy and the possibility of being discovered by detection tools.

The measure is global due to the technical impossibility of limiting tracking geographically, meaning it will affect all countries where Claude is available, including Argentina. This has a direct impact on the local academic and labor spheres, where generative AI use is expanding.

What to keep an eye on

It is important to monitor the implementation of the detection tool and the public verification API that Anthropic will launch. Also, keep an eye on updates regarding the transition period for older models and potential vulnerabilities that may arise. For those who rely on Claude in contexts where AI use is prohibited, it will be crucial to evaluate alternatives or adjust workflows before December.

Practical impact: immediate, conditional, and uncertain consequences

• Immediate: Claude’s new models already include the watermark; texts generated now are detectable.
• Immediate: Users using Claude for work with zero-AI policies risk being detected if the verification tool launches soon.
• Conditional: Starting in December, pre-existing tools will also incorporate the watermark; models released before August 2, 2026, will have a transition period.
• Conditional: Detection may fail in short texts, code, or factual content, which could reduce its effectiveness in certain uses.
• Uncertain: The effectiveness of the watermark against comprehensive rewrites is not yet proven; it could be removed with substantial edits.
• Uncertain: The impact on the ecosystem of developers and companies using the Claude API has not yet been quantified.

Value module: comparison of the watermark with other measures

According to the primary source (official Anthropic statement, cited by La Derecha Diario), the watermark is permanent and irreversible. The SynthID-Text technology was published in Nature in 2024. No official data is available on the detection percentage, but the verification tool will calculate a high probability. No currency conversions or additional calculations are made because no comparative numerical data is provided in the sources.