The system alters the natural randomness of word selection by introducing a statistical pattern imperceptible to the human eye, but easily detectable through a mathematical key.
Claude will launch an invisible watermark to detect AI-generated texts
The measure that Anthropic will implement globally has raised concerns among the millions of users who are returning to use alternatives like ChatGPT or Grok
The system alters the natural randomness of word selection by introducing a statistical pattern imperceptible to the human eye, but easily detectable through a mathematical key.
This digital signature "lives" in the sequence of emitted tokens and survives intact when copied, pasted, or exported across the various integrated platforms of the company's ecosystem, including Claude API, Claude Code, Claude Cowork, and Claude Tag, as well as in services offered on corporate clouds like AWS, Google Cloud, or Microsoft Foundry.
Even visual files and other compatible formats like PNG, JPG, or SVG will be affected, as they will incorporate cryptographic metadata signed under the standard of the Coalition for Content Provenance and Authenticity (C2PA). The invisible watermark will drastically affect translations, although the system presents vulnerabilities in extremely brief texts, in programming code generation, or in strictly factual content.
Furthermore, if a user employs the platform solely to correct minor punctuation or grammar errors in a document written by a human, the model's intervention may not be sufficient to leave a trace. Although a light edit of the text is insufficient to erase the signal, Anthropic admits that a "complete rewrite" of the content could indeed eliminate it.
The final betrayal for the user community lies in the launch of a binary AI detection tool and a public API for verification. This system will calculate with high probability whether the model participated in the writing, exposing users' work regardless of its prior scope.
In this scenario, the warning for professionals, content creators, and the academic sector is clear: If you use Claude for your university or a zero AI policy project, stop using it right now. With the imminent threat of being discovered, a massive wave of millions of users has decided to abandon the platform to return to the safety of ChatGPT or migrate permanently to watermark-free alternatives like Grok.
In a move that has shaken the foundations of the tech industry, Anthropic has decided to strictly align with the requirements of Article 50(2) of the AI Act of the European Union, a regulatory framework that imposes draconian fines of up to 15 million euros on companies that do not clearly identify content generated by artificial intelligence.
Although the regulation originates in Europe, the tech firm will apply this measure globally in all countries where Claude is available, due to the technical impossibility of limiting tracking geographically. This decision for global control has raised alarms among millions of users who trust the privacy of their texts.
The technical implementation of this system is totalitarian: the company confirmed in an official statement that the measure is "permanent and irreversible".
This marking mechanism is already active for new models and will be mandatory in existing tools after December, with a special transition period for those models launched before August 2, 2026.
The system does not use hidden characters or store personal or conversational information, nor does it alter the intellectual property or legal responsibilities of the generated texts. Instead, the software directly manipulates the process by which the model progressively selects words.
To achieve this, Anthropic has adopted a version of SynthID-Text, the sophisticated watermarking technology developed by Google DeepMind and originally published in 2024 in the prestigious scientific journal Nature