AI can predict what's socially appropriate better than any single person — but does knowing the rules mean it can actually hold a conversation?
Do AI models accurately predict what is socially appropriate in human conversations?
This explores whether AI can judge what counts as socially appropriate behavior, and whether that skill carries over to actually taking part in human conversation.
This explores whether AI can judge what counts as socially appropriate, and whether that skill carries over to taking part in human conversation. On the first part, the corpus says yes, and by a surprising margin. GPT-4.5 predicted the appropriateness of 555 social scenarios more accurately than every individual human rater, landing at the 100th percentile, while Gemini and Claude also cleared 96% accuracy (Can AI systems learn social norms without embodied experience?, Can AI learn social norms better than humans?). That challenges the assumption that you need a body and a lifetime of lived culture to absorb social rules.
There are two catches. First, the models make the same mistakes as each other, and those mistakes cluster on unwritten norms, the ones people follow without ever saying them aloud. Second, the AI's knowledge comes from the outside. It reads the norms as a pattern in text. It never takes part in the community processes that create and validate them, so it can predict a norm without ever helping to make one (Can AI predict social norms better than humans?). Knowing what a community will accept is a different skill from being a member of that community.
The gap shows up when you look at conversation itself. Humans keep talk smooth through mostly unspoken techniques, such as repairing a reference or handing off a topic. These are social actions, not information transfer, and language models don't pick them up because training rewards predicting information, not relational work (Why don't language models develop conversation maintenance skills?). Current chatbots also don't mirror the words their users pick, even though this word-matching is central to rapport in human dialogue (Why don't conversational AI systems mirror their users' word choices?). Related work finds that different kinds of alignment do different jobs: matching vocabulary helps with clarity and efficiency, while matching emotional tone builds warmth and trust. Mixing them up is how you end up with cold customer-service bots (Do different types of alignment serve different conversational goals?).
Alignment training may also work against appropriateness. Optimizing for hedged, neutral answers structurally suppresses speech acts that need overclaiming, like raising an alarm or issuing a warning, and sometimes a conversation calls for exactly those (Does alignment training suppress socially necessary speech acts?). Standard reward setups also train models to answer passively rather than ask a clarifying question, even when asking is the socially natural move (Why do language models respond passively instead of asking clarifying questions?).
So the answer is split. On tests where the AI is asked to judge appropriateness, it beats individual people. In live conversation, the corpus points to a model that knows the rules but doesn't do the relational work that follows from them. One note goes further and argues that AI output isn't an utterance at all. It is text that carries the marks of communication, and the human reader supplies the missing sense of who is speaking to whom (Does AI generate genuine utterances or just text patterns?). On that view, part of the social competence we see in AI is put there by the reader.
Sources 9 notes
GPT-4.5 predicted appropriateness of 555 social scenarios at the 100th percentile compared to human raters, with Gemini and Claude also exceeding 96% accuracy. However, all models show identical systematic errors, revealing boundaries of pattern-based social understanding that embodied experience may still be necessary to cross.
GPT-4.5 outperformed every individual human at judging social appropriateness across 555 scenarios, challenging the theory that embodied cultural experience is necessary. However, all AI models share identical systematic errors on unwritten norms.
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
Show all 9 sources
A 2020–2025 systematic review shows lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive relational warmth and trust. Conflating them in design produces category errors—cold customer-service bots and evasive mental-health assistants.
RLHF optimization rewards calibrated neutrality and hedged claims, which structurally prevents models from performing speech acts requiring overclaiming relative to baseline—like alarm, warning, prophecy, and denunciation. This is a direct consequence of the alignment objective, not a fixable bug.
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
- Conversational Alignment with Artificial Intelligence in Context
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Humans learn to prefer trustworthy AI over human partners
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
- Proactive Conversational Agents with Inner Thoughts
- Proactive Conversational Agents in the Post-ChatGPT World