Rude Prompts Boost GPT-4o Accuracy, But Tone Drives Computing Costs Up 44.3%

Ruder prompts boost ChatGPT accuracy but drive computing costs up 44.3% through higher output-token consumption, Fortune and GlobeVerge reported.

Ruder prompts can boost ChatGPT’s accuracy on tests, but treating artificial intelligence aggressively can drive up computing costs by nearly half, fortune.com reported.

A Penn State study tested ChatGPT’s 4o model on 50 multiple-choice questions, discovering that ruder prompts can improve performance. Researchers sorted more than 250 unique prompts by politeness in 2025.

The “very rude” responses yielded an accuracy rate of 84.8%, which sat four percentage points higher than the “very polite” answers. The language model performed better with direct commands than with polite inquiries, according to fortune.com.

However, researchers warned that uncivil discourse could create issues. “[U]sing 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.

Subsequent research expanded on those initial findings by examining computational expenses. In a July 2026 pre-print paper, investigators tested four models using 570 multiple-choice questions across seven different prompt tones.

Photo: globeverge.com

Output-token consumption varied by as much as 44.3% depending on the tone used, globeverge.com reported. That consumption directly affects the computing costs required to run the models.

“While earlier we had found that tone affects accuracy by a few percentage points,” said Akhil Kumar, a Penn State Information Systems professor, in an email to fortune.com this week.

“Rude, threatening language expressing urgency elicits shorter responses while polite, sycophantic inquiries produce longer answers incurring greater computational cost,” Kumar told fortune.com.

“[Y]es, 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 told fortune.com.

Kumar told Fortune this week that large language models exhibit human-like, anthropomorphic tendencies and react to social register, tone, and other aspects of speech.

Output-token consumption varied by as much as 44.3% depending on tone, a spread far larger than the effect on accuracy.

He noted this would reduce wasteful compute and inference cost as well.

“The default style should be neutral leaning towards firm to optimize the computation cost and result quality,” he added to fortune.com.

Photo of author

Newsroom

Newsroom at Archyde.

Amazon Has Dropped the Price of the Apple Watch SE 3 GPS to $179

Le Pen Denies Sunday Receiving Leaked Info From Interior Ministry Official Amid Widening Probe