1d • General discussion
Understanding AI Watermarks
Some writers are confused about the so-called "watermarks" that companies like Anthropic, OpenAI, and Google are adding to AI-generated text.
There are two types of watermarks being used. One is metadata, which is in the file and can be easily stripped out (for example, by copying and pasting text from one file into a new file). The other "watermark" is the authorial voice or style of the generated text, which can be detected based on word choice, so that if you simply copy and paste (or even retype verbatim) the generated text, it can still be detected because the text hasn't changed.
Re-writing AI-generated text in your own words and your own authorial voice can destroy this so-called "watermark."
Here is Anthropic's explanation: How Claude marks AI-generated content
For a technical explanation, read: A Watermark for Large Language Models
7
11 comments
Understanding AI Watermarks
powered by
skool.com/independent-authors-guild-5490
A place for authors who want to get their books out of their heads and into readers' hands, with publisher and bestselling author Lara Helmling.
Suggested communities
Powered by