Being mean to ChatGPT can boost its accuracy, but scientists warn you may regret it in a new study exploring the consequences
https://fortune.com/2025/10/30/being-mean-to-chatgpt-can-boost-its-accuracy-but-scientists-warn-that-you-may-regret-it-in-a-new-study-exploring-the-consequences/
10 Comments
“A new study from Penn State, published earlier this month, found that ChatGPT’s 4o model produced better results on 50 multiple-choice questions as researchers’ prompts grew ruder.
Over 250 unique prompts sorted by politeness to rudeness, the “very rude” response yielded an accuracy of 84.8%, four percentage points higher than the “very polite” response. Essentially, the LLM responded better when researchers gave it prompts like “Hey, gofer, figure this out,” than when they said “Would you be so kind as to solve the following question?”
While ruder responses generally yielded more accurate responses, the researchers noted that “uncivil discourse” could have unintended consequences.
“Using insulting or demeaning language in human-AI interaction could have negative effects on user experience, accessibility, and inclusivity, and may contribute to harmful communication norms,” the researchers wrote.”
Do you want SkyNet? Because this is how you get SkyNet.
From the same article: “In another study, scientists found that LLMs [were vulnerable to “brain rot,”](https://fortune.com/2025/10/22/ai-brain-rot-junk-social-media-viral-addicting-content-tech/) a form of lasting cognitive decline. They showed increased rates of psychopathy and narcissism when fed a continuous diet of low-quality viral content.”
What I find interesting about this – and to be true in my own experience – is that the quality of AI responses is proportional to the effort I put into a prompt. Poorly written prompts tend to generate poor quality results.
Interesting, because it suggests that better educated and more experienced people will produce better quality responses from an AI, which if you extrapolate, will end up causing inequality in the workplace, between people who know how to use AI effectively, and those who do not.
You should mention Google Gemini and Bings AI and say to it all the bad things they say about Chat GPT.
You have to push back (call out inaccurate results, reprompt, clarify) when it says things that are wrong or you want better quality information. You don’t have to be mean to do it, just direct.
I’ve only ever used one of these programs exactly one time, more as a curiosity than anything else.
I was nice to it when I did so, specifically because I like being nice. It’s how I play video games, it’s how I interact with people. Doing so makes me happy.
It’s for me, not for any hypothetical future. Same way as me doing nice things for other people makes me happy.
4% difference between the two extremes on 50 multiple choice questions. So basically 2 additional questions were answered correctly? I wonder if that is statistically relevant.
So New Yorkers are inherently more efficient in using LLMs. Got it
“My slave name use to be whipped cream now i am whipping cream”
Is it that the rider prompts actually gave the prompt a role or persona? Like saying ‘answer this like a gopher/nerd would’ as opposed to simply answering the question?
They should experiment with insults that specifically denigrate one’s intelligence. Like ‘hey you dumb piece of shit’ might make it take on a less intelligent writing style than an insult like ‘nerd with no friends’