Rudeness or politeness: how the tone of a prompt affects language model accuracy

The tone of a prompt to a language model can change the probability of a correct answer, but rudeness is not a universal way to improve the quality of the result. The Mind Your Tone study found a specific effect on GPT-4o: very rude prompts achieved higher accuracy than very polite ones. The difference was four percentage points, which is not enough to consider an aggressive style an advantage for other models and tasks.

A demanding prompt is more useful than an insulting one.

How the Mind Your Tone experiment was conducted

Mind Your Tone studied how the politeness or rudeness of a prompt affects the accuracy of a language model’s answer. The authors prepared 50 questions in mathematics, natural sciences, and history. Each question was rewritten in five tones: very polite, polite, neutral, rude, and very rude. This produced a set of 250 prompts.

All versions preserved the original task but changed the pragmatic framing of the request. The researchers compared not different questions, but different ways of presenting the same question to the model. They evaluated responses using GPT-4o and applied paired statistical tests to compare conditions. This design can reveal an association between tone and outcome within the prepared set, but it does not turn the observation into a universal law governing language models.

What results did they get?

In the experiment described, very rude prompts achieved 84,8% accuracy, while very polite ones reached 80,8%. The difference was 4 percentage points in favor of the rudest version. The authors also reported that impolite prompts generally outperformed polite ones in their set.

For everyday conclusions, the figures look notable, but scientific interpretation calls for caution. The study involved one model, the questions covered only three subject areas, and each question had pre-prepared wording. It is unknown whether the same result would recur on long analytical tasks, programming, translation, document work, or multi-turn dialogues with several clarifications.

Why four percentage points don’t make rudeness a universal technique

One experiment answers a narrow question: how GPT-4o performed on a specific set of 250 prompts. It does not answer broader questions about all language models, service versions, or user scenarios. The result may have depended on the subject matter, the words the authors used to express irritation, prompt length, presentation order, and other features of the test.

The authors of Mind Your Tone themselves emphasize the limits of the conclusion. Their work does not prove that rudeness always improves accuracy or that users should adopt an aggressive style. A stronger conclusion would require different models, independent task sets, more subject areas, and repeated checks under different generation conditions.

The nature of the metric also matters. Accuracy on a multiple-choice test measures selection of the correct option, but does not necessarily reflect the usefulness of an answer in real work. A good business result involves correctly understanding the task, explaining the solution, following the required format, checking constraints, and avoiding invented details. A rude prompt might incidentally improve one metric while worsening others.

How tone changes an answer without creating new abilities

A language model does not gain new knowledge just because the user writes politely or aggressively. It selects tokens sequentially, guided by the prompt context and generation parameters. Changing the words and tone changes the context and, in turn, the probabilities of different continuations.

The authors of the study use this to explain their cautious interpretation of the result: tone can shift the probabilities of word choices without giving the model new abilities. A rude prompt may sometimes sound more like a command demanding a brief, confident answer. In another case, insults may add noise, change the structure of the task, or prompt a defensive response. The same technique can have different effects depending on the model and the content.

Imagine two prompts for the same task. The first contains a long emotional preamble and an unclear goal. The second directly asks the model to solve the task, show intermediate steps, and check the result. If the second answer is better, the reason is not aggression but clarity of instruction. Users often mistake specificity, constraints, and output format for “strength of tone.”

What Sergey Brin said

Sergey Brin stated that language models sometimes work better when the user threatens them, even going as far as threats of physical violence. He himself acknowledged how unusual and limited this observation was. The statement is consistent with the possibility of a tone effect, but it does not replace a controlled study or establish a rule for all models.

Comparing it with Mind Your Tone shows a similarity in direction, not an equivalence in evidence. The study compared pre-prepared prompt variants and measured GPT-4o’s responses to 50 questions. Brin’s statement described an observation without a complete protocol, control sample, or reproducible metrics presented in the source material. It should be treated as an opinion and personal observation, not as an official description of language models.

Why an aggressive style is harmful to users and teams

The authors of Mind Your Tone do not recommend deliberately using aggressive wording. In their assessment, habitual use may normalize rudeness in human communication. The risk is especially apparent in teams where prompts are copied into work instructions, support templates, and internal knowledge bases.

Insults do not clarify the task. They do not tell the model which facts to treat as given, what format is needed, where verification is required, or what to do when data is missing. In a work process, this style makes it harder to audit prompts, train new employees, and transfer a template to another model.

What if the model does answer better to a rude prompt? The user gets a local improvement on a specific task, but not proof of a reliable method. Before adopting it, compare several tone variants on your own data and evaluate not just correctness, but also completeness, format, consistency, and error rate.

How to demand an accurate answer without insults

A strict instruction and aggression are not the same thing. A demanding prompt sets criteria; it does not demean the other party or the system. It helps the model identify the goal and constraints, whereas emotional pressure adds context whose usefulness has not been confirmed.

• Name the task. Write exactly what needs to be done: solve, compare, check, structure, or shorten.
• Set a quality criterion. Ask the model to state assumptions, check calculations, and separate facts from assumptions.
• Specify the format. State the length, structure, table, or sequence of steps if it affects the result.
• Forbid making things up. Ask the model to say explicitly when data is insufficient rather than filling gaps with guesses.
• Add a verification step. Ask the model to reread its answer and list possible errors or weak points.

A practical prompt might look like this: “Solve the problem, show your calculations, check the result in two ways, and state separately which data you assumed.” This prompt remains polite while setting a stricter working mode for the model. Its advantage comes from the instruction’s structure, not emotional pressure.

How to test the effect of tone on your task

If accuracy is critical, compare variants using your own set of examples. First, prepare identical tasks and change only the tone. Then run neutral, polite, and demanding versions under comparable conditions. Record results using criteria chosen in advance; otherwise, a single successful answer can easily be mistaken for a pattern.

1. Collect tasks that reflect the user’s real work.
2. Create several versions of the same prompt without changing the input data.
3. Evaluate answers using a consistent scale: correctness, completeness, format, and consistency.
4. Repeat the test across different topics and, if possible, across several models.
5. Compare average results and individual failures, not just the best example.

This approach does not require treating aggression as a necessary variable. For practical tasks, it makes more sense to first check the prompt’s clarity, the order of requirements, and whether it includes evaluation criteria. Tone should be treated as one experimental variable, not as a ready-made prompting recipe.

Practical takeaway

Mind Your Tone found a difference between very rude and very polite prompts to GPT-4o: 84,8% versus 80,8%. This is an interesting observation about the effect of phrasing, but it is not proof that rudeness makes a model more intelligent or guarantees a better answer.

A reliable strategy for users is to state the task precisely, make constraints explicit, and check the result. Being demanding helps when it translates into measurable criteria. Insults do not give the model knowledge or replace verification. A single experiment is worth considering, but there is no reason to copy its style without conducting your own tests.

Frequently asked questions about prompt tone

Is it true that rudeness always improves AI answers?

No. The Mind Your Tone study showed this result only on GPT-4o and on a specific set of 250 prompts. The authors do not claim that rudeness offers a universal advantage.

Why might a rude prompt have produced higher accuracy?

Tone changes the context and the probabilities of choosing a continuation. The model may have interpreted the demanding tone, brevity, or structure of the wording differently. This does not mean that it gained new knowledge or abilities.

Can you criticize a language model’s answer?

In the summary of Anthropic’s discussion of its rules included in the review materials, the prohibition is described as targeting abuse, not businesslike error checking and direct requests. This is a paraphrase of the discussion, not a verified quotation from Anthropic’s official policy.

What style should you use at work?

Use a neutral or politely demanding style. State the goal, format, evaluation criteria, and steps to follow. This kind of prompt is easier to repeat, share with colleagues, and evaluate against clear criteria.