When AI Gets Polite Wrong: The Science of Social Norms in Chatbots

There is something off about an AI that is too polite.

You tell a chatbot you are frustrated with a billing error. It responds: "I completely understand your frustration! I would be absolutely delighted to assist you with this today!" The words are technically polite. But the reaction is wrong. The enthusiasm is wrong. The word "delighted" is wrong. You know, immediately and viscerally, that something is broken about this interaction.

Now consider the other end. An AI that is too direct: "Your refund was denied. The return window expired on March 12. Contact your bank if you disagree." Accurate. Completely devoid of social intelligence. It reads like a system error message, not a response from anything that understands what the situation means to you.

Both of these are politeness failures. Not in the everyday sense of "being nice," but in the deeper linguistic sense: failures to calibrate language to the social relationship, the emotional stakes, and the communicative needs of the moment.

Most AI deployments land somewhere in this range. Performatively enthusiastic on one end, bluntly transactional on the other. Neither end serves the user. Both damage trust. And the gap between them is not closed by picking a better tone adjective in the system prompt.

ICX has explored the broader language problem this sits within in the post on why your chatbot has a language problem, not an AI problem. Politeness is one of the most visible dimensions of that problem, and one of the most misunderstood. This post is about why that is, and what to do about it.

What Brown and Levinson Understood That AI Teams Have Forgotten

In 1987, linguists Penelope Brown and Stephen Levinson published a theory that changed how researchers think about conversation. Their politeness theory starts with a simple but powerful concept: every person in a social interaction has a "face," a public self-image they want to protect. Face has two dimensions. Positive face is the desire to be liked, respected, and seen as competent. Negative face is the desire to maintain autonomy and not be imposed upon.

Every utterance in a conversation either supports or threatens face. Asking someone a blunt question threatens their negative face by imposing on their time. Pointing out a mistake threatens their positive face by challenging their competence. Skilled human communicators navigate these moments instinctively. They soften a criticism by framing it as a question. They acknowledge effort before pointing out a gap. They offer options instead of directives.

AI has no natural instinct for this. It has whatever behavior was encoded into its instructions. And in most enterprise deployments, face-threat mitigation is not in the system prompt. Nobody mapped out which types of responses threaten user face, and how the AI should navigate each type.

The result is AI that is socially miscalibrated in consistent, predictable ways. The theory that explains why chatbots feel wrong has existed for nearly forty years. The discipline that could apply it to AI design is only now catching up.

Face-Threatening Acts in Real AI Conversations

Face-threatening acts (FTAs in the linguistics literature) show up in AI conversations in ways that are identifiable once you know what to look for.

The unsolicited correction. A user types their account number with a small formatting error. The AI responds: "That account number format is incorrect. Please enter it as XXXX-XXXX-XXXX." Technically right. Unnecessarily corrective. A socially calibrated response acknowledges the intent and helps without making the user feel they failed: "Let me look that up. It helps to format the account number as XXXX-XXXX-XXXX. Can you give it another try?" The information is the same. The face impact is entirely different.

The competence challenge. A user asks about a policy and was given incorrect information in a previous interaction. The AI responds with the correct policy. No acknowledgment that earlier information was wrong. No recognition that the user is working from a mistake they did not make. The technically accurate response is a face-threat because it implicitly corrects the user without addressing the system failure that created their confusion in the first place.

The over-escalation redirect. A user asks something the AI cannot fully answer. The AI immediately says: "I'm not able to help with that. Please contact our support team." No acknowledgment of what the AI understood. No face-saving for the user. No path offered that respects their time and autonomy. The ICX post on what AI should say instead of "I can't help with that" covers this specific failure mode in depth. The short version: how AI exits a conversation matters as much as how it handles it.

Research from the Nielsen Norman Group on chatbot conversation design consistently finds that users abandon chatbots not primarily because answers are wrong, but because the experience of the conversation feels socially effortful. That feeling traces directly to unmanaged face threats.

The Over-Hedging Problem: When Excessive Politeness Undermines Trust

The opposite failure is just as common, and in some ways more insidious because it feels like good intentions.

In an effort to seem warm and helpful, many AI deployments are trained toward exaggerated politeness: affirmations before every response ("Certainly!", "Great question!"), excessive hedging ("You may want to consider...", "I think that might possibly be..."), and disclaimers that cover adjacent scenarios whether or not they are relevant to the question asked.

This performs politeness while violating its actual function. Brown and Levinson are clear that face-threat mitigation strategies should be proportional to the actual face threat in the moment. When someone asks a simple question, an elaborate politeness ritual before the answer signals not warmth but discomfort. The linguistic equivalent of over-apologizing.

Users read excessive hedging as incompetence or evasion. Forrester research on chatbot adoption consistently identifies qualified, hedged responses as a top driver of user distrust. When the AI says "I believe that should be the case, though I may not have the most current information," users hear: "I don't actually know." That is a face threat to the AI's credibility, which reflects back as a threat to the user's confidence in the entire system.

The calibration question is not "how can we make this AI sound more polite?" It is: "what level of face protection does this moment actually require, and what strategies fit that level?"

A Politeness Calibration Framework for AI Design

Brown and Levinson identify five strategies for managing face threats, from direct (no mitigation) to indirect (maximum mitigation). These map cleanly onto conversational AI design decisions.

Direct, unmitigated responses are appropriate for simple, low-stakes interactions where brevity is what the user needs: confirming a booking, providing a balance, acknowledging receipt of a request. No hedging. No preamble. Get in, answer, get out.

Positive politeness strategies affirm the user's positive face: acknowledging their situation, using inclusive language, showing understanding of the goal behind the question. These belong in interactions with moderate emotional stakes such as a return request, a change to an account, or a complaint that has not escalated.

Negative politeness strategies protect the user's autonomy by offering options, softening directives, and framing requests as requests rather than requirements. These fit moments when the AI needs to ask the user to do something or deliver news they may not want.

High-mitigation strategies apply to the highest-stakes moments: delivering genuinely bad news, acknowledging system failures, managing users who have been through multiple failed interactions. These moments require the most face protection. Most AI deployments have no specific instructions for them at all.

Calibrating AI politeness is design work. It requires mapping what each type of interaction demands, not just setting a default tone.

Building this framework into AI design means categorizing the types of interactions your system handles by their face-threat potential, then specifying the appropriate strategy for each type in the system prompt. It is not about writing warmer copy. It is about making deliberate decisions about social calibration at the architecture level.

The AI content design system post covers how to structure these decisions as governance rather than one-off prompt tweaks. Politeness calibration is one of the most important dimensions to document in that system, and one of the most commonly left out.

Why This Is a Design Decision, Not a Tone Adjective

The difference between an AI that earns trust and one that frustrates users often lives in a space that is hard to name but easy to feel. Interactions that feel off. Responses that are accurate but somehow wrong. Politeness that reads as fake. Directness that reads as rude.

Most AI teams address this by adjusting tone language in the system prompt. Add "empathetic." Change "professional" to "warm." These adjustments do not fix the underlying issue because tone adjectives do not encode social strategy. They describe a register without explaining how to navigate the specific social dynamics the AI will encounter.

The distinction between a chatbot and a truly conversational AI is partly about this. ICX explored that boundary in the post on chatbots versus conversational AI. Conversational AI adapts to context, including social context, not just informational context. An AI that can adjust its response to what the user said, but not to the social stakes of how they said it, is still operating from a script.

As AI moves from answering questions to taking actions on behalf of users, these stakes only increase. Harvard Business Review's research on AI in customer service identifies trust as the primary driver of AI adoption among end users, and trust is built interaction by interaction through language that feels socially intelligent. ICX covered the broader landscape of this shift in the post on why AI agents are replacing chatbots in CX: the more autonomous the AI, the more consequential every conversational decision becomes.

Brown and Levinson did the hard theoretical work in 1987. The question now is whether AI teams are willing to apply it. The organizations that are building this level of intentionality into their language layer are creating experiences that feel fundamentally different from the ones that treat politeness as a checkbox. Users notice. They return. They trust.

If you want to evaluate the politeness architecture of your own AI deployment and identify where the social calibration is breaking down, the services page covers how ICX approaches this work. And if you want to start a conversation about what this looks like for your specific context, the contact page is the right place to reach out.

ICX is building a newsletter for CX leaders who want this kind of thinking on a regular basis. It is coming. Stay tuned, and bookmark the blog so you do not miss what is next.