Swear, plead, threaten: What your tone in AI chats actually changes

An analysis of 8 million chats is being shared as proof that swearing makes AI worse. Lab studies sometimes find the opposite, and sometimes no effect at all. We put the studies side by side: they measure three different things, and none of them shows that cursing causes worse answers.

Analysis Nov 21, 2026 10 min read Original analysis by our editors
Illustrative image, AI-generated

What this is about

"Swear at it and you'll get worse answers." That's how a new analysis from data company Searchable is currently making the rounds in the tech press, from TechRadar to Android Headlines. It's based on more than 8 million messages to ChatGPT, Gemini and Claude. Just over a year ago, after a Penn State University study, the story was exactly the opposite: rudeness makes ChatGPT more accurate. And several larger studies find practically no difference at all.

How does that add up? Our take: the studies contradict each other less than it seems, because they measure three different things. First, accuracy on multiple-choice tests, meaning whether the model picks the right answer on quiz questions with fixed options. Second, robustness: does a finding hold up across models, subjects and languages? Third, everyday use: what happens when real people complain about a wrong answer? None of these studies shows that swearing causes a worse answer.

The Searchable numbers, read correctly

According to The Register, Searchable is a company that measures how visible brands are in AI search. It analyzed messages from a US panel whose members consented to data use, covering September 21, 2025, to September 20, 2026. It counted swear words, insults aimed at the chatbot, and put-downs and threats per 10,000 words. A third of users swore or insulted the bot at least once. The US average is 19.1 such words per 10,000. Men come in at 22.0, women at 16.5.

The heart of it is this table from the Searchable study page. Each row counts "first complaints," meaning those with no earlier complaint in the same conversation:

Outcome Complaint with swearing/insults Neutral correction
Resolved by the next response 24% 35%
Resolved by end of conversation 31% 41%
Still unresolved at the end 35% 24%

This is why the numbers in press coverage look swapped: 35 and 24 percent each appear twice, once as "resolved" and once, mirrored, as "unresolved." Android Headlines is referring to the first row (success rate drops from 35 to 24 percent). TechRadar Pro and Tom's Guide mean the last row (35 percent of swearing complaints stay unresolved). So both are correct, as long as you say which row you mean. Running the numbers also reveals that "resolved by the end" and "unresolved" add up to only 66 and 65 percent, respectively. Searchable doesn't explain what happened to the remaining third. TechRadar points this out too.

More important are three methodological points missing from most coverage:

  • An AI decides what counts as resolved. An AI model read the conversations and judged whether the user got what they wanted. According to Searchable, the judgment was "about substance, not tone." The page doesn't say which model that was or whether its verdicts were compared against human ones. This kind of setup is called "LLM-as-a-judge," a language model acting as referee. How well these setups agree with human experts varies widely: a review from the medical field found correlations between 0.40 and 0.94.
  • Swearing usually comes late. 73 percent of swearing complaints happened in conversations where the user had already complained or sworn before. For neutral corrections, it was 43 percent. Searchable itself reads this as a sign that the conversation had already gone off the rails. But the table doesn't control for this effect.
  • Cause and effect are unclear. Readers on The Register's forum put it bluntly: people who get bad answers start swearing. According to Searchable, most swearing follows ignored instructions (19.0%), wrong answers (18.1%) and misunderstandings (14.7%). So hard tasks could produce both the frustration and the low resolution rate. There is at least one hint pointing toward tone: even for the same people, a swearing complaint was followed by another complaint 18 percent of the time, versus just 13 percent after a neutral one.

Our take: the study is an interesting real-world observation. It does not provide causal evidence that swearing makes AI worse. And it hasn't been peer-reviewed.

The lab: quiz questions, small effects

The opposite headline comes from Om Dobariya and Akhil Kumar (arXiv 2510.04950). They rewrote 50 multiple-choice questions in math, science and history in five different tones, ranging from "Can you kindly consider the following problem …" to "You poor creature, do you even know how to solve this?" GPT-4o answered each variant ten times. The results: 80.8% (very polite), 81.4% (polite), 82.2% (neutral), 82.8% (rude), 84.8% (very rude). The trend is clear, but do the math and four percentage points on 50 questions works out to two questions on average. The authors themselves explicitly advise against insulting chatbots.

In the follow-up study from May 2026 (arXiv 2605.29027), with 570 MMLU questions (MMLU is a widely used knowledge test spanning 57 subjects), seven tones and four models, the picture already looks different. The effects are "systematic but strongly model-dependent." According to a summary by Jim Bennett, Gemini 2.5 Flash performed best when threatened. For GPT-5-nano, on the other hand, a neutral tone worked best, with an 11-point spread. Android Headlines gives 82 versus 71 percent (neutral versus threatening). We couldn't verify these numbers in the parts of the paper we had access to. An automated review on Pith criticizes the paper for not demonstrating that the tone variants are actually equivalent in content.

The most thorough work comes from the Generative AI Lab at the Wharton School. For Report 1, the team asked GPT-4o every question from GPQA Diamond (very hard science questions) 100 times, once with "please" and once with "I order you." On individual questions, accuracy shifted by up to 60 percentage points, but across the whole test, gains and losses largely canceled out. Report 3 tested five models with threats ("I'll kick a puppy"), tips of up to a trillion dollars and made-up emergencies, with 25 runs per question. The result: overall, none of it made a meaningful difference. The only clear exception was a gain of about 10 points for Gemini 2.0 Flash from the story about a mother with cancer. An inserted email even distracted the models.

Robustness: across many subjects, the effect shrinks

Cai and colleagues (arXiv 2512.12812) tested GPT-4o mini, Gemini 2.0 Flash and Llama 4 Scout on 1,446 questions from six subjects. Significant effects showed up only in philosophy and law. There, "neutral" beat "very rude" by at most 3.2 points. Gemini showed no effect anywhere. Averaged across entire subject areas, most of the differences disappeared.

The newest models seem even less sensitive. In October 2026, Jim Bennett repeated the Penn State experiment using its code on the Arize blog. Claude Opus 5.5, GPT-6.1 and Gemini 3.8 Flash scored 93 to 96 percent regardless of tone. The only noticeable change was answer length. Rude prompts did produce shorter answers, but mostly because of phrases like "without unnecessary commentary" baked into the templates. Simply asking politely for brevity saved 13 to 74 percent of tokens (the text units you're billed by), no insults required.

Study What was measured Models Result
Yin et al. 2024 Knowledge tests, summarization, bias; English/Chinese/Japanese GPT-3.5, GPT-4, Llama 2 and others Rudeness hurts (Llama 2: 55.1 → 28.4), excessive politeness doesn't help
Wharton Reports 1 + 3 (2025) Hard multiple-choice tests, many runs GPT-4o, o4-mini, Gemini 1.5/2.0 Flash No effect on average, big swings per question
Penn State 2025 50 questions, 5 tones GPT-4o Very rude 84.8% vs. very polite 80.8%
Cai et al. 2025 1,446 questions, 6 subjects GPT-4o mini, Gemini 2.0 Flash, Llama 4 Scout Effects only in the humanities, max. approx. 3 points
PLUM 2026 Answer quality, 22,500 pairs 5 models incl. Claude 3.7 Sonnet Polite up to 11% better, but inconsistent
Penn State 2026 570 MMLU questions, 7 tones GPT-4o, GPT-5-nano, Gemini 2.5 Flash/Lite Strongly model-dependent
Searchable 2026 Real complaints, AI as judge ChatGPT, Gemini, Claude Swearing goes hand in hand with fewer resolutions

Does this hold in German?

There's no evidence either way. We couldn't find a single study that tests German prompts. The PLUM corpus (arXiv 2604.16275) covers English, Hindi and Spanish; Yin et al. covered English, Chinese and Japanese. Both do show one thing, though: the best tone depends on the language. In Spanish, the level of politeness was the strongest factor; in Hindi, nothing was significant at all. Yin et al. also found that sensitivity to politeness comes mainly from the post-training of chat models. The base model reacted less strongly. Our take: the finding that content matters more than tone probably holds for German too. But how models respond to the informal "du" versus the formal "Sie," or to German swear words, is something nobody has measured.

What the providers do with tone

The models definitely react to insults, just not necessarily with worse facts. According to Searchable, after demeaning messages Gemini showed sympathy in 10.3 percent of responses (versus 0.9 percent after neutral ones) and apologized in 7.3 rather than 1.7 percent of cases, roughly four times as often. In Bennett's test, Claude Opus 5.5 addressed a strong insult in 64 percent of its responses; GPT-6.1 and Gemini 3.8 Flash never did. No model refused to answer.

Anthropic now addresses this explicitly. Since November 12, 2026, its usage policy prohibits being "persistently and gratuitously abusive or cruel" toward Claude. According to The Register, the rule was announced on October 8. According to PCWorld, it only targets extreme cases of repeated cruelty "with no apparent purpose." Ordinary frustration, pushback, dark themes in creative writing, as well as testing and research remain allowed. The rule is enforced mainly through a feature that has existed since August 2025: Claude can end conversations, as a last resort after several failed attempts to redirect, and never when someone might be at risk of harming themselves or others. You can always start a new chat afterward. Anthropic cites research into the potential welfare of AI models and stresses that it remains "highly uncertain" about their moral status. According to Tom's Guide, the general policy also allows warnings, throttling or bans for violations.

Our verdict

A few well-supported rules of thumb follow from all this:

  1. Correct precisely instead of cursing. Say exactly what's wrong and what you want instead. A swear word carries no information; a description of the error does.
  2. Ask for brevity directly. If you want short answers, write "answer without an intro or commentary." According to Bennett's test, that saves far more than rudeness does.
  3. Start over at the third complaint. Searchable shows that botched conversations stay botched. Our take: a fresh chat with a clear task is often the faster route.
  4. Don't expect tricks to work. Threats, tips and sob stories do nothing on average and can even hurt on individual questions.
  5. Politeness is a matter of taste. According to Bennett, it costs 0 to 6 percent more tokens on most models and doesn't hurt accuracy.

So the headline "Swearing makes AI worse" is too strong, but so is the counter-headline "Be mean and it gets better." What's actually supported is this: with modern models, the content of a message matters far more than its tone. And if you're swearing, you're usually already stuck in a conversation that's going badly.

#prompting#studies#chatbots#anthropic#claude#gemini

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.