Claude is reported to hide a statistical mark in its text, not in any characters

What is being reported

A Russian-language post on dev.to says Anthropic made a change this summer that you cannot see in the Claude interface or in any answer. Text from newer Claude models now carries a mark. It is not a signature at the end and it is not a string of hidden characters.

This rests on one secondary source, and we have not seen Anthropic's own pages. The post says the company published a help article and, three days later, a blog post explaining the text mark, around 10 and 11 August. Treat the details below as the post's account of those documents.

According to the post, the marking covers all text from Claude models released after 2 August 2026, in the chat app, the API, Claude Code and other products. Older models are to be added later, with no date given. The post adds that Amazon Bedrock, Amazon's cloud service for AI models, should finish its rollout by 10 November.

How a mark can hide in word choice

A language model builds text one piece at a time. At each step it has a list of possible next words with probabilities, and it picks one with a random draw. Usually that draw is plain chance.

The watermark replaces that chance with something steered. A secret key, combined with the few words just before, nudges the model slightly toward some of the acceptable options. The meaning stays the same, but over a long passage the choices lean in a pattern that ordinary randomness would not produce. Someone with the key can detect the lean. To everyone else the text looks normal.

The post says Anthropic builds on SynthID-Text, a method Google DeepMind described in Nature in 2024 and released as open code. Google tested its version on 20 million Gemini responses and, as the post tells it, found no drop in user ratings for marked answers. Anthropic makes the same claim about quality, but the post notes nobody has independently tested it on Claude.

One consequence is that facts and code carry a weaker mark. When there is one right answer, such as a date or a function name, there is little room to vary the wording.

Why now, and who can read it

The post links the move to Article 50 of the EU's AI law, which requires AI-generated content to be recognizable by machines. A voluntary transparency code began applying in August, and the post lists Anthropic, Google, Meta, Microsoft and OpenAI among its signers. It says the marking works worldwide, because Anthropic cannot yet limit it to one region.

Reading the text mark is restricted. The detector is a private preview that people can apply for. The post names regulators, law enforcement, fact-checkers and media, independent researchers, schools and universities, EU civil-society groups, and companies checking their own Claude output. Accuracy, false-positive rate and the minimum text length are not published, only a warning that very short passages do not give enough signal.

Files work differently. Images, video and audio from Claude's code sandbox get signed metadata under a standard called C2PA, recording that Claude made or handled them. A Claude Content Checker on claude.com reads this in the browser, but it does not examine text. PDFs, plain text files and office documents get no signature.

What the mark can and cannot show

The most important limit comes from Anthropic's own wording, as the post relays it: a found mark means content may have been created or processed by Claude, and the signal is not final. Processed is the key word. If you write an article and ask Claude to fix the commas or translate it, marked passages can appear in your text, and you are still the author.

The reverse also holds. No mark does not mean a human wrote it. The text could come from an older Claude model, another AI system that does not mark its output, or an open model. Someone may also have changed it enough to weaken the signal.

Anthropic says the mark survives copy and paste and "may persist" after light edits. Heavy editing, rewording, translation, mixing with other text and cutting down to a couple of sentences weaken it, because the statistical lean is spread across the whole text. For images, re-saving, recompressing or a screenshot drops the metadata.

The post also cites Wired, from 19 August, saying bypass tools appeared within hours of the announcement and one gained thousands of stars on GitHub. Nobody has independently verified whether they work. We are not publishing removal steps. We cannot verify them, and whether to strip a provenance mark is a question for your employer, school or publisher.

What it changes in practice

This is not an AI-text detector in the usual sense. It is a maker's signature that reliably catches an unedited Claude answer and says little about text that passed through a person, a translator or another model.

For students, a found mark is not proof of cheating, since the same mark appears when a student asks Claude to review their own writing. A conversation fits better than a verdict. For bloggers and copywriters, a Claude answer published nearly untouched is probably signed, though few can check today and the post says no platforms or search engines have announced that they read the mark.

For developers, any Claude text your product passes on is marked. The post says Anthropic advises working out which Article 50 duties apply to your own service, because the model's mark does not discharge them for you. The post notes that OpenAI reportedly built a similar mark for ChatGPT years ago and held it back, in part over fears it would unfairly hit non-native English speakers who use the tool as an editor. The post cites The Wall Street Journal for that. The underlying problem is the same.