Hello,

I’m not sure if I actually like the .md file format but it was the easiest way to get started with my portfolio website. Maybe with claude code, I could dive in later to improve upon it.

I started reading SoK: Watermarking for AI-Generated Content today morning as part of my Responsible use of AI & Algorithms course. I have to later fill in a reading response and maybe I will include the responses here too.

The number one thing I struggle with when reading a paper is to think about it critically. Hopefully writing and reading the paper simultaneously will help me ask better questions and engage not just on a superficial level. If you have any tips which might help, let me know!

Thoughts -

  1. Aren’t open-source models a limitation to watermarking ?
  2. They mention improved AI models’ quality surpasses human-generated content, i’m confused in what way, like statistically pixel-wise, what is good quality in that ?
  3. Watermarking for fraud detection is new, I had not come across this before.
  4. What are the pros and cons of an imperceptible watermark ? First thought, based on the assumption that majority people don’t sue/ or have resources to sue, can still be influenced by GenAI content if watermark is invisible to naked AI. Ofc it makes sense from an IP standpoint but policies need to consider more scenarios ?
  5. The industry developments, are they deployed in real time systems or used when requested or disputed ? Are they effecient enough to be deployed 24x7 without the need of a conflict ?
  6. I’m not sure if the paper is going to talk about how to add watermark but more about it as a concept.