|
Actions:
|
2025-01-14 10:28 AEST by Arthur Barrett - Google are working on an open standard to detect AI generated text.
Also images and video, see H4VAPP enh bug Bug 7882 "Support for AI digital watermark detection / SynthID"
https://deepmind.google/technologies/synthid/
SynthID: Tools for watermarking and detecting LLM-generated Text
https://ai.google.dev/responsible/docs/safeguards/synthid
You can read the paper in Nature for a more complete technical description of the method.
https://www.nature.com/articles/s41586-024-08025-4
Watermarking AI-generated text and video with SynthID
https://deepmind.google/discover/blog/watermarking-ai-generated-text-and-video-with-synthid/
SynthID Text
This repository provides a reference implementation of the SynthID Text watermarking and detection
capabilities for the research paper published in Nature. It is not intended for production use. The core library
is distributed on PyPI for easy installation in the Python Notebook example, which demonstrates how to apply
these tools with the Gemma and GPT-2 models.
https://github.com/google-deepmind/synthid-text
Introducing SynthID Text
https://huggingface.co/blog/synthid-text
SynthID must be used to watermark the text. At the point in time I'm not sure how many publicly available
models actually do this. Does Gemini even do it 'by default'? All this was just released in late 2024 (maybe:
October 23, 2024?). So adoption may grow. The SynthID for images seems to be used widely on Google
already after a couple of years, and has entered the consciousness of general users (to click on 'about this
image').
How easy is it to break? It wasn't hard to find an example:
https://gist.github.com/natzir/76e690fd4f5826f3746a109ad08cff16?short_path=62f2144
In terms of implementation - I think this would be server side, but I could see a use case for clients to ask this
question before committing. But I think the primary use case would be to prevent commit of AI generated
content?
On the server, a plugin like 'audit' to implement rules would be ideal? see bug 7883: "enh: server plugin to
enforce checkin rules" |