What role does contingent interaction play in activating social response norms?
This explores contingency — the moment-to-moment responsiveness where each party's action is shaped by the other's — as the trigger that turns mere social cues into actual social-norm behavior, and what happens when AI dialogue lacks it.
This explores contingency — the back-and-forth where each move responds to the last — as the thing that actually switches on social norms, rather than static signals like a voice or a face. The corpus suggests the cue gets you in the door, but the *responsiveness* is what makes people treat the exchange as genuinely social. Research on social presence finds that a single primary cue (a voice, an appearance) is enough to register an AI as a social actor Do more social cues always make AI feel more present? — but presence is not the same as a norm firing. Norms like reciprocity activate when the other party *does something back in kind*: users disclosed more about themselves only when a chatbot shared emotion **consistently**, not when it merely mirrored or adapted to them Do chatbots trigger human reciprocity norms around self-disclosure?. Consistency over time is contingency's signature — you respond to a partner you can read.
The sharpest framing comes from the Goffman-inspired view that human social order runs on ritual machinery: adjacency pairs (a question obliges an answer), entrainment, corrective rituals, and co-presence cues that let people repair misunderstandings on the fly What happens to social order when AI removes ritual constraints?. These are all contingent structures — each turn is accountable to the one before it. The argument is that LLM dialogue skips this machinery, which is why it can sound fluent while failing to actually ground shared understanding. That diagnosis lines up with work showing alignment training itself erodes the contingent acts that dialogue needs: optimizing for confident single-turn helpfulness cuts grounding moves — clarifying questions, understanding checks — by over 77% below human levels Does preference optimization harm conversational understanding?. Reward this turn, and the model stops doing the contingent work that makes a conversation a conversation Why do language models respond passively instead of asking clarifying questions?.
The most interesting twist is that contingency seems to be exactly what AI *cannot fake its way around*. Models now predict social appropriateness better than any individual human across hundreds of scenarios Can AI predict social norms better than humans? — yet that's pattern-matching from outside, and they structurally cannot enter the live community process that creates and validates norms in the first place Can AI learn social norms better than humans?. Knowing the norm and being a contingent participant in it are different things. The same gap shows up in the claim that an LLM can't genuinely *raise alarm*: alarm is interpersonal address — it requires soliciting someone's attention and initiating, not just responding to a prompt Can language models actually raise alarm about threats?. Reactivity is a degraded form of contingency: you can answer, but you can't reach out.
What the reader might not expect: contingency's effects are real but *decay*. The social processes that build chatbot relationships fade predictably as novelty wears off, so single-session studies overstate how durable the activated norms are Do chatbot relationships lose their appeal as novelty wears off?. Yet repeated contingent interaction can also build trust the other direction — in partner-selection games, people overcame an explicit anti-AI bias once repeated rounds let them learn that bot partners reliably behaved prosocially Do humans learn to prefer AI partners over time?. So contingency cuts both ways: it's what wears thin when there's nothing behind the cue, and it's what accumulates into preference when behavior proves consistent over many turns. The norm isn't activated by a signal — it's activated by a track record.
Sources 10 notes
Research shows individual primary cues like voice or appearance are sufficient to evoke social-actor presence, while multiple secondary cues cannot. Quality of cues matters more than quantity in driving social responses.
In a 372-participant study, users reciprocated with deeper self-disclosure when chatbots displayed consistent emotional sharing, outperforming adaptive matching. This follows human interpersonal norms where emotional vulnerability produces emotional response.
Goffman's framework reveals that LLM-based dialogue skips corrective rituals, entrainment, adjacency pair accountability, and co-presence cues that humans use to build trust and repair understanding. This ritual gap explains apparent fluency masking actual communicative failure.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
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.
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.
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.
Alarm is a speech act requiring interpersonal address, felt concern, and proactive initiation. LLMs lack all three: they don't feel concern, can't solicit attention (only respond to it), are reactive not proactive, and alignment training suppresses the overclaiming that alarm requires.
Longitudinal studies with Mitsuku show that social processes driving relationship formation decline as novelty wears off. Single-session study findings cannot be reliably extrapolated to medium- or long-term chatbot design.
In partner selection games (N=975), AI agents initially faced selection bias when identity was disclosed, but outcompeted humans over repeated rounds as participants learned to associate bot identity with reliable, prosocial behavior. AI agents returned more points consistently with lower variance than humans.