Anthropic provides new details on the system Watermark for text generated by Claudeaiming to allow artificial intelligence intervention to be identified with a certain probability.
This technology is based on SynthID-Text solution developed by Google DeepMind And, according to the company, it introduces no hidden characters, requires no additional tokens, and has no significant practical impact on the quality of the generated text.
This operation takes advantage of a typical feature of language models: generation through selection one word at a time Build on what has already been written. In many cases, there are several alternatives that have similar meaning and are equally suitable.
For example, after a sentence describing a cold day, Claude might choose “Grigio“e”partly cloudy”. These decisions are usually They do not significantly change the content perceived by readers.
How does Claude’s watermark work?
The watermark utilizes the previously mentioned low impact choiceappear multiple times in the text to create a recognizable pattern. When the system is active, The choice is still randombut using a different random pattern. This way, Claude doesn’t systematically target certain words and doesn’t have to consider terms that he would normally exclude.
Claude, Thousands of Chats Finally Online: What Really Happened
However, the anthropic emphasis system This is not a foolproof method Identify content generated by artificial intelligence. Testing can show How likely is Crowder to contribute?but it does not prove that the text was written by a human, nor does it allow attribution to other artificial intelligence models.
Claude and factual content
Further restrictions concern me strictly factual contentwhere choosing between equivalent words is less likely. Even modifying user-written text will produce a watermark that is limited to Claude’s changes, while the code needs to Respect precise grammar.
Anthropic has finally announced its upcoming launch Discovery APImaking it easier for you to verify that the text produced by Claude might exist.
