SynthID: Identifying AI-Generated Content | Hacker News Reader

SynthID: Identifying AI-Generated Content

deepmind.google

24 pointstosh7 comments

Watermarking scheme for traitor tracing.

... by watermarking the generated content.

... with a proprietary algorithm, the details of which we will not say a word about

They are providing this tool for users to run. Eventually someone will just use this tool en mass, then use this data to create a local model that will detect stuff with Google's watermark.

> They are providing this tool for users to run

Where? From what I see in https://deepmind.google/technologies/synthid/#:~:text=Where%... , it is not a tool, it is an "undescribed algorithm". Otherwise people will immediately:

1) build an algorithm which slightly modifies few pixels to break SynthID.

2) apply SynthID to real images to imitate tampering made "allegedly" by specific people

I can do better than that:

def is_ai(text):

  return text.contains(["crucial", "pivotal", "craft", "multifaceted"])

“as an AI assistant”

Reply on news.ycombinator.com