I was interested to see this on the heels of the discussion about tipping/threatening LLMs a few days ago [1] (discussed here: [2]).
Based on the examples at the end of the paper, it looks like their inquiry was somewhat orthogonal to the tip/threat investigation. That is, the tips and threats can be done in a way that is polite, when "rude" is defined as calling the LLM a scumbag. Putting these two studies together, I wonder if the best outcome could be achieved by politely asking the LLM to do XYZ, offering a tip if it succeeds, and noting that terrible things will happen if it fails.
Based on the examples at the end of the paper, it looks like their inquiry was somewhat orthogonal to the tip/threat investigation. That is, the tips and threats can be done in a way that is polite, when "rude" is defined as calling the LLM a scumbag. Putting these two studies together, I wonder if the best outcome could be achieved by politely asking the LLM to do XYZ, offering a tip if it succeeds, and noting that terrible things will happen if it fails.
1: https://minimaxir.com/2024/02/chatgpt-tips-analysis/
2: https://news.ycombinator.com/item?id=39495476