You can remove an AI watermark from text by asking a second AI

The EU AI Act AI outputs to be "detectable as artificially generated" and "embedded within the content in a manner that is difficult for it to be separated from the content." 

But Sean Goedecke, a software engineer, explains why the European Union's new watermark rule for AI text won't work. 

Text watermarking hides the signature in words it chooses. The model favors words that score highly on a simple formula, and a high total score across a block of text indicates that a machine wrote it.

This kind of watermarking is easy to remove. Goedecke says you can strip it by asking any weak, unwatermarked model to paraphrase the text. "Because the watermark is inherent to subtle vocabulary choices, re-wording the content will remove the watermark," he says. 

The Act also requires watermark methods to be interoperable, so the AI companies will have to publish how theirs work.

Previously:

• California makes AI companies leave fingerprints on their slop
• A $5 AI book draft that fools writing detectors