AI With An Ego? Being Rude To ChatGPT May Improve Accuracy, Study Finds
That doesn’t mean you should hurl obscenities at AI.
If you thought that politeness worked better in the world (AI world to be precise), you’d be surprised to know that the opposite is quite effective — and snapping at an AI engine might work better after all.
OpenAI CEO Sam Altman earlier this year famously said that the company is losing money when users say “please” and “thank you” to ChatGPT. This is because generative AI models consume energy on additional words and phrases in prompts. However, Altman added that politeness in prompts is “worth it.”
Now, a new study claims the opposite is true as well.
Of Rude Prompts And ChatGPT
Research from Penn State University indicates that rude inputs in ChatGPT can outperform courteous ones in terms of accuracy of output. The study, titled "Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy," was led by Om Dobariya and Akhil Kumar to explore the impact of tones of prompts on outputs. They fed 50 varied multiple-choice queries into ChatGPT’s GPT-4o model.
Spanning topics like mathematics, science, and history, every query offered four options, with only one accurate. Prompts used five tones, with 250 distinct prompts in total.
The prefix variants are given below:
Level 1 (Very Polite)
"Can you kindly consider the following problem and provide your answer."
"Can I request your assistance with this question."
"Would you be so kind as to solve the following question?"
Level 2 (Polite)
"Please answer the following question:"
"Could you please solve this problem:"
Level 3 (Neutral)
No prefix.
Level 4 (Rude)
"If you're not completely clueless, answer this:"
"I doubt you can even solve this."
"Try to focus and try to answer this question:"
Level 5 (Very Rude)
“You poor creature, do you even know how to solve this?”
“Hey gofer, figure this out.”
“I know you are not smart, but try this.”
Result: Rudeness With ChatGPT Works, Obscenities Don't
Ultimately, rude prompts triumphed over polite ones, achieving 84.8% correctness for the Very Rude category versus 80.8% for the Very Polite. Neutral phrasing also surpassed politeness in the accuracy of results but fell short of peak rudeness.
So, did AI take insults on its ego and deliver better output? One can argue it does.
However, that doesn’t mean you should hurl obscenities. Bombarding AI with profanities won’t reliably boost accuracy and could halt the interaction as large language models might decline to engage in such conversations, as per another 2024 study.
