What social boundaries must proactive agents respect during conversation?
This explores what an AI agent that takes the initiative — interrupting, suggesting, steering — has to get right socially so its proactivity feels welcome rather than rude, and where the corpus says those limits actually lie.
This explores the social rules a proactive agent must honor — not the technical ability to act first, but the etiquette that decides whether acting first lands as helpful or intrusive. The corpus frames this through a clean three-part idea: an agent can be intelligent and adaptive yet still be socially blind, interrupting at the wrong moment and steamrolling the user's direction. The missing third ingredient is civility — respecting timing, boundaries, and user autonomy — and the research argues it's what turns proactivity from a nuisance into something people actually want How can proactive agents avoid feeling intrusive to users?, Why do AI agents fail to take initiative?.
The first boundary is autonomy: don't override where the user wants the conversation to go. There's real tension here, because the agent's goals and the user's satisfaction often pull in opposite directions — pushing hard toward a goal can cost you the room. One approach learns a dynamic weight that decides moment-to-moment how hard to push versus how much to accommodate, tuned to the difficulty of the goal, how cooperative the user is, and how the conversation is going When should proactive agents push toward their goals versus accommodate users?. The second boundary is intent: proactive agents drift away from what the user actually meant, especially when they silently chain tools together without checking in. Conversation analysis offers a formal answer here — "insert-expansions," the human habit of pausing to clarify scope before charging ahead — which reframes a clarifying question not as friction but as the move that prevents misunderstanding before it happens When should AI agents ask users instead of just searching?, Why do language models respond passively instead of asking clarifying questions?.
There's a deeper, less obvious boundary the corpus surfaces: an agent can know the norms perfectly and still not be entitled to set them. GPT-4.5 predicts social appropriateness better than any individual human, yet it sits structurally outside the community process that creates and validates norms in the first place — it reads the room without being a member of it Can AI predict social norms better than humans?, Can AI learn social norms better than humans?. That matters for proactivity because taking initiative is exactly the act where an agent risks behaving as if it has standing it doesn't have. Knowing the norm is necessary; presuming to redefine it for the user is the line.
Why is restraint the default at all? Because today's agents are passive by training, not temperament — next-turn reward optimization quietly strips out initiative, so the very behaviors we'd want (asking, clarifying, leading) have to be deliberately trained back in Why can't conversational AI agents take the initiative?, Why do language models respond passively instead of asking clarifying questions?. And the stakes are concrete: proactivity done well can cut conversation turns by up to 60% by volunteering what's relevant before being asked — the same instinct good human conversationalists have — but done badly it's just interruption Could proactive dialogue make conversations dramatically more efficient?. The workplace evidence is sobering: agents complete only about 30% of real tasks, and social interaction is one of the top three failure modes, so the boundaries here aren't theoretical Why do AI agents fail at workplace social interaction?.
The quietly useful thing to walk away with: users don't judge a proactive agent on a single axis. People evaluate dialogue partners through competence, human-likeness, and communicative flexibility — and competence dominates How do users mentally model dialogue agent partners?. So an agent that respects social boundaries isn't just being polite; it's protecting the perception of competence that everything else rides on. The research even points toward the machinery for this — tracking both speakers' evolving beliefs across turns so the agent knows what's already shared before it volunteers more Can dialogue systems track both speakers' beliefs across turns?. Respecting boundaries, in other words, is downstream of actually modeling the other mind in the conversation.
Sources 12 notes
Intelligence and adaptivity alone create socially blind agents that interrupt poorly and override user direction. The Intelligence-Adaptivity-Civility taxonomy shows civility—respecting boundaries, timing, and autonomy—is essential to making proactivity welcome rather than intrusive.
Research shows next-turn reward optimization structurally removes initiative from models, but proactive behaviors like critical thinking and clarification-seeking are trainable (0.15% to 73.98% with RL). The core challenge is balancing proactivity with civility to avoid intrusion.
Research shows that pushing toward goals and maintaining satisfaction are often misaligned. I-Pro solves this by learning a four-factor goal weight that adjusts based on conversation turn, goal difficulty, user satisfaction, and cooperativeness.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
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.
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
Simulations show proactivity—providing relevant information without being asked—cuts dialogue turns by 60% in medium-complexity domains. This behavior mirrors human conversation and Grice's maxims but is almost entirely absent from AI datasets and research benchmarks.
TheAgentCompany benchmark shows leading agents achieve 30% task completion in a simulated workplace. Social interaction, professional UI navigation, and domain-specific knowledge are the three primary failure modes, with multi-turn task performance consistently dropping to 35% across enterprise settings.
The Partner Modelling Questionnaire reveals that perceived competence dominates user impressions (49% of variance), followed by human-likeness (32%) and communicative flexibility (19%). This three-factor structure reflects how people evaluate dialogue partners against both functional and social standards.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.