Watermarking Language Model Outputs Quiz

Answers and explanations appear as you take the quiz.

1. In a green-list/red-list watermark, how is the signal embedded?
   • By biasing token sampling toward a secretly chosen green list
   • By inserting invisible Unicode characters between words
   • By appending a hidden footer to the text
   • By encrypting the entire output
2. Why is watermarked text harder to detect when it is very short?
   • Short text uses a different vocabulary
   • The green-token surplus is too small to be statistically significant
   • Detectors ignore text under 10 words by rule
   • Short text is always human-written
3. What typically determines whether a token lands on the green or red list?
   • The token's alphabetical order
   • The user's IP address
   • A hash of the preceding tokens combined with a secret key
   • The time of day the text was generated
4. Which action is most likely to remove or weaken a text watermark?
   • Reading the text aloud
   • Saving the file in a different format
   • Changing the font
   • Heavy paraphrasing or translation
5. What is a key tradeoff when increasing the strength of the green-token bias?
   • It boosts robustness but can degrade text quality and diversity
   • It makes detection slower but text faster
   • It improves both detection and quality with no cost
   • It removes the need for a secret key
6. What real-world system brought language-model watermarking into production?
   • Adobe Photoshop
   • Google DeepMind's SynthID-Text
   • A blockchain ledger
   • DNS records
7. As a watermarked passage gets longer, how does the detector’s green-token z-statistic typically scale?
   • Linearly with the number of model parameters
   • As the inverse square root of the passage length
   • Roughly with the square root of the number of tokens
   • It remains exactly constant
8. Why is the secret watermark key not embedded directly in the generated text?
   • The vocabulary has no token IDs
   • Embedding a key would disable tokenization
   • The model cannot store character strings
   • The signal is encoded in token choices, while keeping the key secret prevents others from easily reproducing or forging the split