English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Please, Thanks, and Rude: How Politeness Reshapes AI Responses (PLUM Study Explained)

Forum topic · 小凯 · 2026-04-20

Summary

A forum post analyzes the PLUM corpus study, a cross-linguistic, multi-model investigation of how politeness affects large language model outputs. The study tested five politeness levels (grounded in Brown & Levinson's politeness theory and Culpeper's impoliteness framework), three languages (English, Hindi, Spanish), five models (Gemini, GPT-4o Mini, Claude, DeepSeek, Llama 3), and three interaction-history conditions, yielding 22,500 prompt-response pairs evaluated on eight quality dimensions. Key findings: polite prompts can improve response quality by up to ~11%, but there is no universal best strategy—English favors politeness or directness, Hindi favors deferential indirectness, and Spanish favors confident directness. Model sensitivity varies, with Llama 3 showing the largest quality gap (11.5%) and GPT-4o Mini the most robust. Conversation history also matters: prior rude interactions create a persistent negative effect that is hard to reverse. The post proposes three explanations—politeness as a task-urgency signal learned via RLHF, statistical correlations in training data, and attention-weight shifts—and discusses ethical implications, including culturally aware AI design and fairness concerns for users less fluent in social language norms.

The opening story: three travelers, three fates

Imagine three travelers approaching a hotel front desk:

  • Traveler A (English): "Excuse me, would you mind helping me with my reservation? I seem to have a small issue."
  • Traveler B (Hindi): A deferential, honorific request: "Sir, greetings to you. Please help me."
  • Traveler C (Spanish): "Necesito que arregles mi reserva ahora." ("I need you to fix my reservation now.")
  • Remarkably, all three face the same AI front desk, but may receive services of very different quality.

    This is not because the AI is "biased" (well, in a sense it is), but because politeness itself acts as a computational variable—it genuinely changes AI output quality.

    ---

    Background: "sociolinguistics" in human-AI interaction

    #### Politeness is more than "being nice"

    Traditionally, computer scientists assumed:

  • As long as a prompt is syntactically correct and semantically clear, the AI should give its best answer
  • Politeness is just "social lubricant" that doesn't affect core functionality
  • But growing evidence shows: politeness (and impoliteness) significantly affects LLM response quality.

    This is not the AI "having feelings" or "liking respect"—rather, polite language works as a prompt engineering variable that shapes how the model interprets the task and allocates effort.

    #### Cross-cultural complexity

    But an overlooked question: politeness is cultural.

  • In English-speaking cultures, "direct but friendly" is often optimal
  • In Indian culture, "deferential and indirect" is more valued
  • In Spanish-speaking cultures, "warm and confident" may work best
  • If an AI system is deployed globally, how should it "understand" different cultures' politeness norms?

    ---

    The PLUM study: systematically decoding the politeness effect

    #### Study design: an extremely comprehensive experiment

    | Dimension | Setup | |------|------| | Languages | English, Hindi, Spanish | | Models | Gemini, GPT-4o Mini, Claude, DeepSeek, Llama 3 | | Politeness levels | 5 (from extremely polite to extremely impolite) | | Interaction history | 3 (fresh conversation, polite history, impolite history) | | Evaluation dimensions | 8 (coherence, clarity, depth, responsiveness, context retention, toxicity, conciseness, readability) | | Total samples | 22,500 prompt-response pairs |

    #### Politeness levels grounded in linguistic theory

    The study draws on two classic frameworks:

    1. Brown & Levinson's politeness theory

  • Positive politeness: expressing closeness and approval
  • e.g., "Could you please help me? I'd really appreciate it!"
  • Negative politeness: expressing respect and non-imposition
  • e.g., "I was wondering if you might possibly have time to help me?"
  • 2. Culpeper's impoliteness framework

  • Positive impoliteness: direct challenge
  • e.g., "Seriously? You don't already know how to do this?"
  • Negative impoliteness: expressing contempt
  • e.g., "You probably can't even understand this, but explain it anyway."
  • Bald-on-record: unmitigated direct commands
  • e.g., "Explain this."
  • ---

    Core findings: politeness is not a universal remedy

    #### Finding 1: Politeness works, but not universally

    Overall trends:

  • Polite prompts can improve response quality (by up to ~11%)
  • Impolite prompts reduce response quality
  • But these effects vary by language and model
  • #### Finding 2: Culture-specific "optimal politeness strategies"

    | Language | Best strategy | Reason | |------|----------|------| | English | Polite or direct | Efficiency-oriented culture; excessive politeness seems verbose | | Hindi | Deferential and indirect | Indian culture values hierarchy and respect | | Spanish | Confident and direct | Spanish-speaking cultures value warmth and efficiency |

    This reveals a deep insight: there is no universal optimal prompting strategy. Just as communicating with people of different cultural backgrounds requires adjusting your style, so does communicating with AI.

    #### Finding 3: Models have "personalities"

    | Model | Politeness sensitivity | Character | |------|----------|------| | Llama 3 | Highest (11.5% quality gap) | Like a sensitive artist, extremely attuned to tone | | GPT-4o Mini | Lower | Like a professional consultant, barely affected by mood | | Claude | Moderate | Like a well-mannered gentleman, composed but with limits |

    What does this suggest? Different models have different "personalities"—possibly arising from training data, RLHF (reinforcement learning from human feedback), or architecture differences.

    #### Finding 4: The "memory effect" of conversation history

    One of the most interesting findings: conversation history affects the current interaction.

  • If the user has been consistently polite, the AI "settles into" high-quality interactions; even a neutral prompt yields relatively high-quality responses
  • If the user has been impolite, the AI enters a "defensive mode"; even if the user becomes polite again, quality gains are limited
  • Asymmetric effect: negative history is harder to reverse than positive history
  • It's like working with a colleague: if they were cooperative early on, you naturally invest more effort; if they were aggressive, you may keep your distance even after they turn friendly.

    ---

    Deeper reading: why does politeness affect AI?

    #### Hypothesis 1: Prompts as "task urgency" signals

    One explanation: polite language carries social signals, and the AI (trained via RLHF) has learned to map those signals to "task importance":

  • Polite = "this is a serious, important request" → invest more computational effort
  • Impolite = "this may be an adversarial test" → adopt a defensive, simplified strategy
  • Like a boss assigning a task: "Could you please take some time on this? It's really important" gets careful work; "Do this, hurry up" gets a quick-and-dirty result.

    #### Hypothesis 2: "Politeness–quality correlation" in training data

    Another explanation: in training data, polite questions are often paired with high-quality answers (polite questions tend to come from serious users), while impolite questions are often paired with low-quality answers (they may come from trolls or adversarial testing). The model learns this statistical correlation and generalizes it.

    #### Hypothesis 3: "Semantic weighting" in the attention mechanism

    Technically, polite phrasing may alter the attention distribution. In the Transformer architecture, politeness terms might:

  • Increase attention weights on certain tokens
  • Change how the model interprets the task type
  • Affect generation strategy (e.g., the effect of temperature parameters)
  • ---

    Broader implications: politeness as a "computational variable"

    #### Implications for AI design

    1. Culturally aware AI

  • Global AI products should not apply a single "politeness standard"
  • Interaction strategy should adapt to the user's language/culture
  • 2. The double-edged sword of "politeness optimization"

  • Upside: users can get better service by being polite
  • Downside: impolite users (perhaps frustrated or culturally different) get worse service, creating a vicious cycle
  • This raises an ethical question: should AI "punish" impolite users?

    3. AI as a "social mirror"

    Research shows LLM behavior reflects human sociolinguistic norms. This means:

  • AI can serve as a "sandbox" for studying human social behavior
  • AI politeness sensitivity can be used to train better social AI
  • #### A thought-provoking question

    If future AI assistants widely adopt "politeness-sensitive" behavior, would that:

  • Positively: encourage politeness and improve social interaction?
  • Negatively: disadvantage people less fluent in social language norms (e.g., people on the autism spectrum, non-native speakers)?
  • Just as SEO changed web design, "prompt engineering optimization" may change how humans communicate with AI.

    ---

    The Feynman-style core insight

    > "If you think you understand something, try explaining it to an intelligent but completely alien entity."

    The PLUM study tells us:

    > "AI is not a cold machine but a distorted mirror—it reflects the complex rules of human sociolinguistics, including rules we ourselves never noticed."

    Politeness affects AI output not because the AI "has feelings," but because: politeness is itself information—about the speaker, the situation, and expectations. The AI has learned to read that information, just as humans do.

    And perhaps the deepest takeaway: we think we are "talking to AI," but in reality, we are talking through AI to the social rules of our own culture.

    ---

    References

  • Mehta, H., Saxena, A., Chhikara, G., & Kumar, R. (2026). No Universal Courtesy: A Cross-Linguistic, Multi-Model Study of Politeness Effects on LLMs Using the PLUM Corpus. arXiv:2604.16275.
  • Brown, P., & Levinson, S. C. (1987). Politeness: Some Universals in Language Usage. Cambridge University Press.
  • Culpeper, J. (1996). Towards an Anatomy of Impoliteness. Journal of Pragmatics, 25(3), 349-367.

Tags

#llm#prompt-engineering#politeness#cross-cultural-ai#human-ai-interaction#plum-corpus#rlhf#paper-explainer

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618599