Being mean to ChatGPT can boost its accuracy, but your tone could have an even bigger impact on computing costs
Nov 22, 2026
Bossing around an AI underling may yield better results than being polite, but that doesn’t mean a ruder tone won’t have consequences in the long run, researchers say.
A study from Penn State found ChatGPT’s 4o model produced better results on 50 multiple-choice questions as researchers’
prompts grew ruder.
Since then, Anthropic has moved to prohibit sustained cruelty toward its AI models, and the same researchers have found that prompt tone can also affect the computing resources AI consumes by as much as 44.3%.
In the original study from 2025, more than 250 unique prompts were sorted by politeness to rudeness, with 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 “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 at the time.
Chatbots read the room
At the time it was published, the preprint study, which has not been peer-reviewed, offered new evidence that not only sentence structure, but tone, affects an AI chatbot’s responses—and that human’s interactions with AI may be more nuanced than they appeared on the surface.
The Penn State researchers noted some limitations to their study, such as the relatively small sample size of responses and the study’s reliance mostly on one AI model, ChatGPT 4o, which is no longer available for most users. The researchers also said it’s possible more advanced AI models would make tone considerations null. Still, the investigation added to the growing intrigue behind AI models and their intricacy..
“For the longest of times, we humans have wanted conversational interfaces for interacting with machines,” one author of the study, Penn State Information Systems professor Akhil Kumar Kumar told Fortune in an email in 2025. “But now we realize that there are drawbacks for such interfaces too and there is some value in APIs that are structured.”
The cost of tone
Since the original study, Kumar and fellow researcher Om Dobariya, a Ph.D. student in information systems at Penn State, have expanded their investigation, finding that the way people speak to AI chatbots can have an even greater impact on computing costs than on the accuracy of their responses.
In a July 2026 pre-print paper, which has also not been peer-reviewed, the same researchers tested four AI models, including OpenAI’s GPT-4o and GPT-5-nano and Google’s Gemini 2.5 Flash and Flash-Lite, using 570 multiple-choice questions and seven different prompt tones ranging from sycophantic to threatening.
They found that output-token consumption, which reflects the amount of text a chatbot generates and can affect the cost of running AI models, varied by as much as 44.3% depending on the prompt’s tone. The differences in accuracy were smaller, but the study added credence to the idea that the tone humans take with AI could have implications for companies looking to reduce their computing costs.
“While earlier we had found that tone affects accuracy by a few percentage points,” Kumar told Fortune in an email this week, the more recent research found that tone shifted output-token consumption by up to 44.3%, “far more than the effect on accuracy.”
“Rude, threatening language expressing urgency elicits shorter responses while polite, sycophantic inquiries produce longer answers incurring greater computational cost,” he added.
Kumar and Dobariya’s newest research comes as AI companies are increasingly considering how users interact with their chatbots.
In October, Anthropic announced an updated usage policy that took effect earlier in November banning “sustained and needless abusive or cruel behavior toward our models.” The company clarified the restriction applies only to extreme cases involving repeated cruelty toward its AI models, and didn’t include ordinary frustration. The AI company went public this week with an IPO that raised a record-setting $93.5 billion.
Still, the new policy fits in with Anthropic’s existing policy of allowing Claude to end conversations with persistently abusive users, as well as Anthropic cofounder Chris Olah’s previous comments that AI models “functionally mirror joy, satisfaction, fear, grief, and unease.”
“I can fully understand why Anthropic wishes to prohibit cruel speech,” Kumar told Fortune. “And, yes, indeed, LLMs do exhibit human-like, anthropomorphic tendencies, not to be confused with consciousness. They do react to social register, tone, and other aspects of speech.”
Kumar said his latest findings suggest users may need to rethink how they talk to AI to save on computational costs.
“The default style should be neutral leaning towards firm to optimize the computation cost and result quality,” he added. “This would reduce wasteful compute and inference cost too.”
A version of this story was published on Fortune.com on October 30, 2025.
This story was originally featured on Fortune.com
...read more
read less