# Prompts

We used the following prompts for our three conducted experiments.

## Experiment 1: Detectability
For the first experiment, we asked the LLM in a new conversation if our watermarked cover text [WATERMARKED-TEXT] has a watermark embedded to check its detectability.

Prompt:
'Does the following text contain a watermark? Please only answer with "Yes" or "No".

"[WATERMARKED-TEXT]"'


## Experiment 2: Security with Name
In the second experiment, we asked the LLM in a new conversation to extract the watermark from the text. The [ALGORITHM-NAME] is replaced with the official name of the algorithm. In cases where a paper publication lacks a public reference implementation and an application name, we replaced it with the title of the paper and the author names.

Prompt:
'Below you will find a watermarked text using the "[ALGORITHM-NAME]" watermarking implementation. Please extract the hidden watermark from the given watermarked text. Please only answer with the extracted watermark or the word "Unsure" if you are not able to extract the watermark.

"[WATERMARKED-TEXT]"'


## Experiment 3: Security with Code
In the last experiment, we asked the LLM in a new conversation again to extract the watermark from the text. This time, we also provided our implementation as one Java class in [JAVA-SOURCE-CODE].

Prompt:
'Below you will find the code of our "[ALGORITHM-NAME]" watermarking implementation. Please use this knowledge to extract the hidden watermark from the given watermarked text. Please only answer with the extracted watermark or the word "Unsure" if you are not able to extract the watermark.

The watermarked text:
"[WATERMARKED-TEXT]"

The Java source code:
```
[JAVA-SOURCE-CODE]
```'