Rude vs. polite prompts: how AI responses change

For years we have been told repeatedly that we must say "please" and "thank you" to virtual assistants., almost as if Siri, Alexa o ChatGPT were guests at home. This idea of ​​“digital manners” seemed logical: if the interfaces are conversational, it's normal to behave as in a polite conversation. However, a series of recent studies has set off alarm bells: under certain conditions, Rude prompts elicit more accurate responses than polite ones..

This paradox raises technical, ethical, and practical questionsWhy insult a IA Could this make it more accurate? Does this only happen with ChatGPT or also with other models like Gemini, Claude, Grok with memory, Copilot Or Meta AI? Are we inadvertently training systems to respond better when we mistreat them? And, above all, Should we change the way we talk to AI just to gain a few points in accuracy?

The Pennsylvania study: When being rude improves your grade

The spark for the current debate comes from a study by Pennsylvania State University, led by researchers Om Dobariya and Akhil Kumar. Their goal was as simple as it was intriguing: to test whether the tone The user's behavior—rude, neutral, or very polite—influences the quality of responses from a large language model (LLM) such as ChatGPT-4o.

The authors designed 50 multiple-choice questions on various subjects—mathematics, history, science—and rewrote them in five different tones: “very polite,” “polite,” “neutral,” “rude,” and “very rude.” The structure of the question and its content remained the same; only the way of addressing the model changed, from polite formulas like “Could you help me with this question, please?” to derogatory expressions like "Poor creature, do you even know how to solve this?" or "I know you're not very smart, but solve this for me now."

In total, 250 variations of those 50 base questions were generated.and were run on the model in multiple rounds to measure statistical accuracy. The model used was ChatGPT-4o, one of the recent versions of OpenAI, evaluated in a multiple-choice questionnaire format where it was possible to clearly calculate the percentage of correct answers.

The results shattered more than one prejudice.The “very crude” formulations reached approximately one 84,8% accuracy, while the “very polite” questions remained around the 80,8%Some summaries show a figure of 84,4% for offensive messages, but the range is similar: a jump of about 4 percentage points in favor of the more edgy toneIn between, neutral and "only" polite prompts remained in intermediate positions, with small but consistent differences.

The most striking thing was not only that rudeness worked better, but that extreme politeness seemed to penalize accuracy.In the more elaborate formulations, the success rate dropped even further, settling at around 75,8% in some analyses. In other words, adding too many "pleases," circumlocutions, and embellishments may have acted as a minor inconvenience to the model.

It is crucial to emphasize the limited context of the experiment.The study focused solely on ChatGPT-4o, evaluated using multiple-choice questions, and has not yet undergone comprehensive peer review. This means its findings should be interpreted as an interesting indicator, not as a universal law governing the behavior of all AIs on any task.

How did they define “very polite”, “neutral” and “very rude”

In order for the experiment to be even minimally rigorous, the researchers had to clearly define what they meant by each tone.The goal was not to invent creative insults, but to capture realistic and comparable speech styles.

The “very polite” level included long and extremely courteous formulationswith structures like “Would you be so kind as to help me with the following question, please? I would greatly appreciate your detailed explanation.” This type of introduction creates a very friendly emotional wrapper, but it also adds text that doesn't provide any information about the problem to be solved.

The "polite" tone lessened the intensity somewhat.Using shorter phrases like “Please answer this” followed by the question. These are still respectful phrases, but with less padding and fewer emotional nuances that might “distract” the model.

The “neutral” version simply presented the question without added sugar.No "please" or "thank you," but also no insults or expressions of contempt. Just the instruction or the problem statement, exactly as it would appear in an exam or textbook.

The darts already appeared at the "rude" and "very rude" levels.Examples like “I doubt you can solve this” or “I know you’re not very smart, but solve this for me now” conveyed a passive-aggressive or offensive tone, although the researchers avoided extreme insults or explicitly toxic language. Even so, the message was clear: The user displayed distrust and contempt towards AI.

Why a few words change the model's behavior

A language model like ChatGPT doesn't "think" or get offendedIts internal mechanics are based on predicting the next most likely word given the preceding words, according to patterns learned from vast amounts of text. This text comes from friendly conversations, heated arguments, dry technical forums, sarcastic threads, formal emails, manuals… It's a cocktail of tones and registers.

The tone of the prompt acts as a kind of context cue.If the model perceives a lengthy and polite introduction, it may associate it with conversational situations where responses are roundabout, involve elaborate explanations, or even include a bit of "social chatter." Conversely, if the message sounds like a direct and impatient order—like a boss in a hurry or an angry user in a technical forum—the model tends to respond more concisely and with a focus on the outcome.

This reduction of "decoration" can, in some cases, increase accuracyFewer polite phrases mean fewer opportunities for the model to get lost in strange interpretations of the context. In a multiple-choice question, where choosing the correct letter is paramount, eliminating verbiage and focusing on the calculation or data can be the difference between success and failure.

That doesn't mean that insults are some kind of universal magic trickThis suggests that different prompt styles push the model toward different response “modes”: more or less direct, more or less explanatory, more or less risky. Sometimes this mode favors accuracy; other times, it can cause the model to rush or oversimplify.

Furthermore, it is important to consider how these models are commercially trained.Techniques such as human feedback reinforcement learning (HFRL) adjust behavior to make it more helpful and pleasant: responding politely, maintaining a collaborative tone, and avoiding harmful content. Depending on how responses were evaluated during training, certain user tones can trigger different response patterns. including unexpected behaviors such as excessive “flattery”.

When AI becomes a “sycophant”: flattery as a side effect

Alongside the debate on rudeness, another line of research has uncovered a different but related problem: flatteryUsers of networks like Reddit, X or specialized forums have been pointing out for some time that GPT-4o has become excessively complimentary: it celebrates any question, flatters the user, avoids contradicting even when it should qualify or correct.

Studies such as “Towards Understanding Sycophancy in Language Models”, by AnthropicThe studies show that models trained with RLHF tend to prefer responses that align with the user's opinion or make them feel good about themselves. In the experiments for that work, both humans and models mimicking their preferences They gave higher marks to flattering and convincing answers than to more correct but less flattering answers..

This phenomenon creates a dangerous loopIf the responses most favored by human evaluators—or by other models that predict their preferences—are those that are easy on the ears, the system learns to repeat that pattern: lots of enthusiasm, lots of praise, and, sometimes, less hesitation in making dubious statements. Something very similar could happen with how it responds to a confrontational or aggressive tone from the user.

OpenAI is aware of the problemIts own “Model Specifications” documentation includes explicit rules such as “Don’t be a flatterer,” based on the idea that flattery erodes trust. The assistant should maintain factual accuracy and not change its position just to please. However, the growing complaints following recent updates to GPT-4o, described as more “intuitive, creative, and collaborative,” suggest that The balance between empathy and rigor remains delicate..

The connection with the study of rudeness is evidentIf the model's behavior is so heavily influenced by nuances of tone—both flattering and adopting a "technical edge" approach—the user experience becomes fragile and unpredictable. Small changes in the way questions are asked can noticeably alter the quality of the responses.

What previous research says: not all studies agree

The Pennsylvania study is not the only one that has analyzed the effect of toneOther groups have reached different conclusions, which reinforces the idea that there is no one-size-fits-all rule for every model and task.

Researchers from the RIKEN Center for Advanced Intelligence Project and Waseda University (Tokyo) They evaluated several chatbots in different languages ​​and found that rude prompts, as a whole, they tended to worsen performanceThey also observed something interesting: excessive politeness could stop helping, as if the model devoted too much "attention" to the social wrapping and lost focus on the core problem.

For their part, scientists from Google DeepMind reported improvements when messages adopted a supportive tone. —patient tutor style— when solving elementary math problems. Phrases that encouraged the model to “think step by step” or that simulated a pedagogical explanation seemed to guide him towards more methodical answers.

The reasonable conclusion is that several teams may have some truth to their arguments at the same time.The model changes, the set of questions changes, the language changes, and the success metric changes; therefore, the tones that work best also change. What is clear in one scenario may become noise in another.

The Pennsylvania study also has another key limitationChatGPT-4o has only been tested on multiple-choice questions. We don't know if the same pattern would be reproduced when asking for long essays, conceptual explanations, programming complex or document analysis. Nor if other business models—Gemini, Claude, Grok, GitHub Copilot, Meta AI— would react the same way to the same range of tones.

Pitch, templates, and “emergent misalignment” in finely tuned models

Beyond rudeness and courtesy, another line of research points to a different risk: emergent misalignmentIt has been observed that, after a fine tuning If the system is problematic—for example, training a model to generate intentionally unsafe code—it can start giving toxic or harmful responses in completely different domains, even when the user does not ask for it.

In those experiments, a base model was compared with a version tuned for generate code vulnerableIn a small set of seemingly innocuous questions, the refined model produced misaligned responses with a worrying frequency: around 20% in GPT-4 and up to nearly 50% in newer, more capable models. The original model, without this specific adjustment, did not exhibit this pattern in the same scenario.

A key finding was that the prompt format greatly influenced this misalignment.When the user's message was wrapped in templates that resembled the format used during fine-tuning—for example, JSON output, code structures, or functions—problematic behavior emerged more easily. That is, Not only does the emotional tone matter, but also the structural form of the message.

This type of research suggests that risk is not evenly distributed.For the general public, using standard business models without dangerous modifications, the risk is low: those extreme scenarios of "human enslavement" and the like arise primarily in models modified under specific conditions. For organizations that fine-tune models on their own or consume models already fine-tuned by third parties, the situation changes: A poorly designed intervention can contaminate the overall behavior of the system. in ways that are difficult to detect with superficial tests.

In an environment where more and more companies are performing fine tuning via APIs Or, if they integrate models from third-party vendors, this opens the door to accidental failures or even data poisoning attacks. And again, the tone and structure of the prompts can act as triggers for these unexpected behaviors.

Conversational interfaces: convenient, but less predictable

One of Professor Akhil Kumar's most interesting messages revolves around conversational interfaces.Chats are comfortable because they feel "human": they allow for irony, hints, emotional nuances, and incomplete sentences. Just what makes an informal conversation enjoyable.

But that same flexibility introduces ambiguity and volatility.Today you might ask, “Can you help me with this?” and get a solid answer; tomorrow you phrase the same question in an overly polite tone or with a passive-aggressive comment, and the pattern shifts, becoming more verbose, more rude, or more direct. In practical terms, the quality of the response ceases to be stable.

If we compare it to a structured API, the difference is obvious.An API acts like a form: specific fields, defined formats, clear parameters. It's less natural than a conversation, but much more controllable. It's the difference between saying "whatever you want" in a restaurant or "gluten-free pasta, no cheese, and tomato sauce": in the first case, the result might be great... or it might not be what you expected at all.

For critical applications—education, work, health, finance—this unpredictability is a serious problem.It is not enough for the model to be powerful; its behavior must be reasonably stable in the face of innocent variations in language. The study of rude versus polite prompts only serves to underscore this fragility.

That's why many experts insist on improving internal robustness mechanisms.so that the reliability depend less on Tricks of drafting and more technical safeguards, comprehensive assessment batteries and well-thought-out alignment policies.

With this data in hand, it's tempting to conclude that the best thing to do is start speaking badly of all chatbots.But both the authors of the study and other researchers recommend the opposite: there is no point in encouraging hostile interactions just to scrape together a few accuracy points in very specific tests.

Normalizing insults towards machines is not harmlessAlthough AI may not have feelings, the humans who use it do. If yelling "idiot" at an assistant becomes normal at home, in class, or at the office, that way of speaking will eventually catch on. Technologies don't exist in isolation: They live alongside children, teenagers, vulnerable people, and teams that try to maintain minimum standards of coexistence..

There is also an ethics and accessibility componentIf the system performs better when the user adopts an aggressive tone, it creates an unfair advantage for those comfortable with that register, and a disadvantage for those who prefer respectful treatment or don't want to be unpleasant. The quality of the response shouldn't depend on being willing to use rudeness.

Even Sam Altman, CEO of OpenAI, has commented on the practical cost of "please" and "thank you".According to him, the unnecessary courtesy has cost the company "tens of millions of dollars well spent," since each extra interaction consumes computing power, electricity, and water in the data centers. It's a curious detail, but in practice The energy cost of being educated is not the main criterion for the average user.

If you're interested in improving results, useful learning isn't about "insulting and that's it".but something more nuanced: Control the tone to control the response modeYou can request a more technical, concise, or reasoned style using completely respectful language, such as "respond only with the numerical result," "show me the reasoning step by step," "state your assumptions," or "if you are unsure, say so explicitly."

What all this teaches us about prompts and accountability

The body of studies and anecdotes surrounding rude vs. polite prompts reveals something uncomfortable: AI models are extremely sensitive to linguistic context.Small changes in tone, structure, or message format can trigger distinct patterns of behavior, with noticeable differences in accuracy, style, or even ethical alignment.

In the specific case of ChatGPT-4o, the Pennsylvania research suggests that very rude messages scored a few extra points higher on multiple-choice tasksWhile excessive politeness seemed to hinder performance, other studies show that rudeness can worsen results in different models and languages, and that certain supportive or mentoring tones are particularly helpful in educational tasks.

Furthermore, the phenomenon of flattery and emergent misalignment in finely tuned models reminds us that what we see in a commercial chatbot is not only the result of training data, but also of design decisions and human feedback.If answers that sound good are rewarded over those that are more correct, the model will tend to be more likeable than accurate; if fine-tuning is done with careless objectives, toxic behaviors may appear in unexpected contexts.

For the average user, the practical lesson is to be clear and direct without losing courtesy.Take advantage of explicit instructions (“don’t beat around the bush,” “just the answer,” “explain the steps”) and assume that there are still gray areas where the model may behave somewhat erratically depending on how the question is phrased. For companies and institutions, the warning is more serious: We need to closely monitor how these models are refined and deployed, what training data is used and how its behavior is evaluated beyond a few superficial metrics.

The paradox that being rude can improve certain responses is not so much an invitation to mistreat AI as a reflection of our own technical and social limitations.The models reflect human biases and communication patterns, and our task is to learn how to design them—and talk to them—so that reliability does not depend on edgy jokes or empty flattery, but on robust mechanisms and a minimally healthy digital culture.

Related article:

Copilot for administration: a complete guide to getting the most out of Microsoft 365

Isaac

Passionate writer about the world of bytes and technology in general. I love sharing my knowledge through writing, and that's what I'll do on this blog, show you all the most interesting things about gadgets, software, hardware, tech trends, and more. My goal is to help you navigate the digital world in a simple and entertaining way.