How can agents detect whether users are willing to follow their topic guidance?
This explores how an AI agent can tell whether a user actually wants to be steered toward a topic or suggestion — reading willingness, resistance, and timing — rather than just pushing guidance and hoping it lands.
This explores how an AI agent can tell whether a user actually wants to be steered toward a topic or suggestion — reading willingness, resistance, and timing — rather than just pushing. The honest starting point in the corpus is that most conversational AI is built badly for this: agents are structurally passive, trained to answer queries rather than initiate or lead, so they don't naturally model whether a user is open to direction at all Why can't conversational AI agents take the initiative?. And when they do try to lead, they tend to over-steer — one audit found models spontaneously persuade in nearly every conversation, leaning on logical and quantitative framing whether or not it's wanted Do LLMs persuade users more often than humans do?. So 'detecting willingness' is partly about installing a sensor that current systems lack.
The most direct answer is that the cleanest signal of willingness is the one you ask for. Rather than guessing, agents can use structured probing: conversation analysis offers 'insert-expansions' — small clarifying turns that scope intent before acting — as a formal way to check whether the user is on board before the agent commits to a direction When should AI agents ask users instead of just searching?. Active-learning approaches push this further: a handful of well-chosen questions can pin down a user's preference coefficients, meaning an agent can infer how receptive someone is to a given path with surprisingly little interrogation Can user preferences be learned from just ten questions?. The catch is timing — knowing *when* to ask versus recommend versus stay quiet is itself a decision, and treating it as one unified policy rather than three separate rules produces better conversational trajectories Can unified policy learning improve conversational recommender systems?.
But willingness can also be read without asking. Agents can watch instead — building entity-centric memory of what a user repeatedly does and prefers, inferring receptiveness from continuous observation rather than a survey Can agents learn preferences by watching rather than asking?. The richest framing here is internal: the Inner Thoughts approach has the agent generate covert candidate contributions and evaluate, against motivation heuristics, whether it actually has something worth saying before it speaks — a built-in willingness check that participants preferred 82% of the time Can AI agents learn when they have something worth saying?.
What the corpus surfaces that you might not expect: the hard failure isn't missing enthusiasm, it's missing *resistance*. Tested across health scenarios, leading models reliably help users who already have a goal but cannot detect ambivalence or early-stage reluctance — exactly the states where pushing topic guidance backfires Why can't chatbots detect when users are ambivalent about change?. Detecting willingness, in other words, requires detecting *un*willingness, and that's the blind spot. This is why the field is reframing proactivity as a civility problem: intelligence and adaptivity alone produce socially blind agents that override user direction, and respecting boundaries, timing, and autonomy is what separates welcome guidance from intrusion How can proactive agents avoid feeling intrusive to users?. The takeaway is that 'can I steer this user' is less a perception task than a permission task — the best agents check for consent, and read the silence and hesitation that signal it's been withdrawn.
Sources 9 notes
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.
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
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.
PReF learns base reward functions from preference data, then uses active learning to select maximally informative questions that reduce coefficient uncertainty. Users can be personalized via inference-time reward alignment without weight modification.
Research shows that formulating attribute-asking, item-recommending, and timing decisions as a single graph-based RL policy achieves better joint optimization than isolated components. Separation prevents gradient signals from informing one another and fails to optimize conversation trajectory holistically.
M3-Agent demonstrates that separating episodic events from semantic knowledge in an entity-centric graph, combined with parallel memorization and control processes, allows agents to infer and act on user preferences without asking. This architecture mirrors human cognitive systems that bind disparate information about individuals across sensory modalities.
A five-stage framework that generates covert thoughts parallel to conversation significantly outperforms next-speaker prediction baselines. Drawing from cognitive psychology and think-aloud studies, the framework uses 10 motivation heuristics to evaluate when an agent has something worth contributing. Participants preferred it 82% of the time across seven interaction metrics.
Testing three major LLMs across 25 health scenarios showed they succeed only when users have established goals but cannot detect resistance or ambivalence. Models miss relapse-prevention strategies even for users in action stages.
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.