Hacker News Reader

An interesting thing that Google does is to watermark the AI generated videos using the [SynthID technology](https://deepmind.google/technologies/synthid/).

It seems that the SynthID is not only for AI generated video but for image, text and audio.

View on HN

I would like a bit more convincing that the text watermark will not be noticeable. AI text already has issues with using certain words to frequently. Messing with the weights seems like it might make the issue worse

Not to mention, when does he get applied? If I am asking an llm to transform some data from one format to another, I don't expect any changes other than the format.

It seems really clever, especially the encoding of a signature into LLM token probability selections. I wonder if synthid will trigger some standarization in the industry. I don't think there's much incentive to tho. Open-source gen AI will still exist. What does google expext to occur? I guess they're just trying to present themselves as 'ethically pursuing AI'.

Reply on news.ycombinator.com