Is Anthropic Quietly Watermarking AI-Generated Text?

The entire X is currently discussing reports that Anthropic has begun introducing invisible watermarks into text generated by its AI models. This is not about metadata or a special symbol that can be spotted by the human eye. The marker is embedded directly into the text: users cannot see it, it is not supposed to change the meaning, style, or quality of the material, and it can survive ordinary copy-and-paste. The system operates at the model level, meaning it could potentially cover content generated by Claude across the app, API, Claude Code, and cloud platforms operated by Amazon, Google, and Microsoft. For images such as .png, .jpg, and .svg, a different mechanism is used: a cryptographic signature based on the open C2PA standard, which is designed to establish the provenance of digital content.

The move is primarily connected to AI regulation. From August 2, 2026, relevant requirements under the European AI Act call for certain AI-generated content to be machine-detectable. That is why the current system reportedly focuses mainly on models launched after that date, while older models remain in a transitional period and their output is not currently marked in the same way. But how can a watermark possibly be hidden inside ordinary words? Anthropic has not disclosed the technical details, but Google’s similar SynthID-Text system is already used with Gemini, giving us a general idea of how such technology can work. Imagine a model completing the sentence: “In the morning, I had a coffee and ___ to work.” It could choose “drove,” “walked,” “headed,” or another natural option. A secret key combined with the surrounding context can pseudorandomly divide possible tokens into two groups, after which the model slightly favors tokens from one of them. The reader sees completely normal text, while a specialized detector analyzes the statistical pattern. If, across a sufficiently large sample, the distribution of certain tokens differs significantly from what would normally be expected, the system can estimate that the text was generated by a particular model. In other words, there is no simple list of “watermarked words” and no visible marker: the watermark is statistical and context-dependent.

That is why conventional AI humanizers and prompts such as “write like a human” may prove largely ineffective. They can change the style and wording without necessarily destroying the underlying statistical pattern. Short texts are also much harder to detect reliably because statistical systems need enough material to produce a meaningful signal. A short post or a few sentences may not provide sufficient evidence, while a long article, research paper, or thesis gives the detector a much larger statistical sample. Code presents another challenge because there is generally less freedom to choose between synonyms and alternative formulations. There is also an important complication: a watermark could potentially appear even when a person originally wrote the text themselves and Claude merely edited, translated, or reworked it. Therefore, the presence of a watermark does not prove that the entire text was written by AI. It may only indicate that a particular model was involved in creating or processing the final version.

The biggest difference between this technology and conventional AI detectors is that it does not try to guess whether a text was generated by looking for a particular writing style. AI can produce perfectly natural prose without obvious “machine-like” expressions and still leave a statistical trace behind. In simple terms, this is not a visible mark hidden between words, but rather a digital fingerprint of the model itself. If Anthropic is using a mechanism similar to SynthID-Text, there is no practical way to identify the watermark by looking at the text, and replacing individual words would not necessarily remove it. A deeper rewrite by another model could theoretically alter the statistical pattern, but even that would not guarantee that the watermark disappears. Anthropic has also not yet released a complete public verification system and is expected to provide more technical documentation later.

This creates a rather fascinating new reality. Until now, AI-generated text was mainly detected through its style: overly polished sentences, repetitive structures, characteristic wording, and predictable patterns. The approach is now changing: AI does not have to look like AI — it only has to leave an invisible statistical trace. If technologies like this become standard, the internet could gradually acquire an additional layer of infrastructure in which humans see ordinary text, images, or video, while machines can see information about their origin. And that raises the biggest question of all: if AI-generated content becomes ubiquitous, who will have the right to verify where digital content came from — and how mandatory will invisible provenance markers eventually become across the internet?