Urgent.News

What's breaking now, across hundreds of outlets.

AI

Being mean to ChatGPT can boost its accuracy, but your tone could have an even bigger impact on computing costs

A newer study by the same researcher finds that prompt tone can affect AI's computing costs by as much as 44.3%.

Being mean to ChatGPT can boost its accuracy, but your tone could have an even bigger impact on computing costs

Dealing harshly with AI systems might lead to better outcomes than using a gentle approach, according to a study conducted by researchers from Penn State. The researchers found that ChatGPT's 4o model produced superior results on 50 multiple-choice questions when prompted with increasingly rude language, as the researchers' demands grew more aggressive.

Since then, Anthropic has taken steps to discourage excessive rudeness towards its AI models, while the same researchers discovered that the tone of a prompt can also impact the amount of computing resources consumed by the AI. In the original 2025 study, over 250 prompts were ranked from polite to very rude, with the "very rude" response yielding an accuracy of 84.8%, which was four percentage points higher than the "very polite" response.

In essence, the language model responded more accurately when given prompts like "Hey, gofer, figure this out" compared to more polite requests such as "Would you be so kind as to solve the following question?" Although ruder prompts generally resulted in more accurate responses, the researchers warned that "uncivil discourse" could have unintended consequences.

They noted that "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." AI interfaces The study, which has not yet undergone peer review, provided fresh evidence that not only sentence structure but also tone affects an AI chatbot's responses.

Human interactions with AI may be more complex than they appear at first glance. In an email to Fortune, one of the study's authors, Akhil Kumar Kumar, an Information Systems professor at Penn State, commented on the research, stating, "For a long time, humans have desired conversational interfaces for interacting with machines.

However, we now realize that there are downsides to such interfaces too, and there is some value in structured APIs." The cost of tone In light of the initial study, Kumar and his research partner, Om Dobariya, a Ph.D. student at Penn State, have expanded their investigation. In a July 2026 preprint paper, also not peer-reviewed, they examined four AI models – OpenAI's GPT-4o and GPT-5-nano, along with Google's Gemini 2.5 Flash and Flash-Lite – using 570 multiple-choice questions and seven different prompt tones ranging from sycophantic to threatening.

The study revealed that output-token consumption, which measures the amount of text generated by a chatbot and can influence the cost of running AI models, varied by up to 44.3% depending on the tone of the prompt. While earlier research had shown that tone can affect accuracy by a few percentage points, this new study added weight to the notion that human behavior towards AI could impact companies' computing expenses.

"While we previously found that tone affects accuracy by a few percentage points," Kumar told Fortune, "this 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 latest research comes as AI companies are increasingly considering how users interact with their chatbots. In October, Anthropic announced an updated usage policy, effective earlier in November, which prohibited "sustained and needless abusive or cruel behavior toward our models." The company clarified that the restriction applies only to extreme cases involving repeated cruelty towards its AI models, and does not include ordinary frustration.

The AI company recently went public, raising a record-setting $93.5 billion through its IPO. However, the new policy aligns with Anthropic's existing policy that allows Claude to end conversations with persistently abusive users, as well as Anthropic co-founder Chris Olah's earlier comments about AI models "functionally mirroring joy, satisfaction, fear, grief, and unease."

Kumar told Fortune that he understands why Anthropic wants to prohibit cruel speech, and acknowledged that LLMs do exhibit human-like, anthropomorphic tendencies. He believes that users may need to reconsider how they communicate with AI to save on computational costs. "The default style should be neutral, leaning towards firm, to optimize computation cost and result quality," he concluded. "This would reduce wasteful compute and inference cost too."

Rewritten by urgent.news; may contain errors — check the original.

Source: fortune.com

Read the original at fortune.com →

Report a problem with this story

More in AI

More from Sunday 22 November →