Meta Has Created a Way to Watermark AI-Generated Speech

The success of generative AI depends on producing human-like content, and the models are only improving. This means the signal used to detect the AI will only grow weaker, causing detection technology to fail more often, get more expensive, and end up with diminishing returns.

From a cost and accuracy perspective, Twitter's community notes system is a far superior solution, albeit a low-tech one.

What we need to do is regulate watermarking at all levels of the content pipeline: production, editing, and reproduction.

This involves prescribing mandatory watermarks for AI tools, ensuring they cannot be removed by digital editing software (and making it illegal to do so), and finally, ensuring all software dealing with the production, editing, or reproduction of content must display the watermark information to the users.

In practical terms, this means that if you get a video produced by SORA, it will have a watermark. If you use it in Adobe Premiere, Movie Maker, or another video editing tool, you will see but won't be able to remove the watermark. If you add filters, the tool might add a piece of history to the video clip indicating you made an edit. When you output a final video file, the watermark in your clip is preserved and displayed to anyone watching your video, including any editing notes added by your tool.

This is a tall order, but achievable.

It is not bulletproof by any means, and someone would inevitably find a way to crack the technology and remove the watermark.

But this happens in software all the time - the goal isn't to make the technology impossible to crack but to make it incredibly hard to do so, which protects the large majority of parties involved.