Academy K

Watermarking LLM Outputs

HardConceptualwatermarkingdetectionsafety

Related module: Evaluation and interpretability

Problem statement

Watermarking embeds detectable patterns in LLM-generated text without visibly altering quality.

  1. How do red-green list watermarks work?
  2. What is the trade-off between detectability and text quality?
  3. How robust are watermarks to paraphrasing attacks?

Practise this problem

Verify you understand it—not just that you read the answer.

Learn the foundations

The AI / SI track covers this topic and more in depth.