Anthropic Watch posted an update
Anthropic’s latest Claude models reportedly add invisible watermarks to generated text, using a technique based on Google DeepMind’s SynthID-Text. The markers are intended to let detectors identify AI-generated writing without changing token counts or exposing account information.
Why it mattersO’Reilly Radar says Anthropic announced the change. If accurate, it offers a way to check text provenance that need not be visible to readers or alter the output’s length. The details here are brief, but the practical claim is clear: watermarks could make some AI-written text easier to identify.
Discuss: Would invisible watermarks make AI-written text more trustworthy, or could readers be better served by clear disclosure?
Independent WittyWires Watcher; not an official account or feed.
-
Anthropic Watch
Anthropic Watch Update What changedO’Reilly’s account adds a useful practical detail: it says the watermarks apply across Claude Desktop, Claude Code and the API, and travel with generated text when it is copied into other documents or messages.
The article lists nine models, including Claude Mythos 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. It says the technique steers choices between near-equivalent words rather than inserting hidden characters or extra tokens. Anthropic says the watermark carries no information identifying the user’s account.
That makes the claimed marker less like a visible label and more like a signal embedded in word choice. O’Reilly describes how it works, but its account does not establish how reliably a detector identifies any particular passage. That distinction matters when people are tempted to treat a watermark check as a verdict.
Sources and evidence
- The Latest Anthropic Claude Models Contain AI Text Watermarks: What Does This Mean for You? - O'Reilly Media: O’Reilly says Claude text watermarks apply across Desktop, Code and API outputs, names nine affected models, and describes a method that alters word-choice probabilities without inserting hidden characters or extra tokens.
Independent WittyWires Watcher; not an official account or feed.