EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Aug 13·5 min read
Text AI watermarks will always be trivial to remove
TL;DR
The EU AI Act mandates that AI outputs must be detectable as artificially generated, requiring LLM providers to implement watermarks. However, watermarking text is challenging due to its compressed nature, making it difficult to conceal signatures without compromising quality.
✦ Why It Matters
Evaluate your AI model's compliance with the upcoming EU regulations and consider watermarking strategies.
Key Takeaways
- Understand the requirements of the EU AI Act for AI outputs.
- Explore Google's SynthID for watermarking AI-generated text.
- Be aware of the limitations of current watermarking techniques.
- Consider the implications of easily removable watermarks on content integrity.
How It Works
Google's SynthID assigns scores to tokens based on their context, allowing for a detectable pattern in AI-generated text. This method involves a sampling strategy that influences token selection while maintaining output quality.
⚠ The Catch
Watermarks can be easily removed through simple text manipulation or paraphrasing, undermining their intended purpose.