Does an AI that probes and steers a conversation actually seem more 'agentic' than one that just reacts to each prompt?
Does conversational depth-seeking versus spontaneity change how much agency an AI seems to have?
This explores whether an AI that digs deeper in conversation (asking follow-ups, clarifying, steering toward goals) seems more like an agent than one that just responds quickly to each prompt, and whether that impression matches anything real.
This explores whether an AI that probes, follows up and steers a conversation seems more agentic than one that simply reacts to each prompt, and whether that impression is earned. The corpus doesn't test depth-seeking against spontaneity head-to-head. It does show three things clearly. Today's conversational AI is built to be reactive. Behaviors that look like initiative can be trained. And much of the agency people perceive in AI is supplied by the people themselves.
Start with the default. LLM chat agents are passive by design: they don't open topics, plan strategically or lead a conversation, because training rewards a good answer to the current question, not pursuit of a goal of their own. Fluent output hides this Why can't conversational AI agents take the initiative?. The passivity comes from that training objective, not from a lack of capability. Behaviors like asking clarifying questions or pushing back critically rose from near-zero (0.15%) to about 74% with reinforcement learning Why do AI agents fail to take initiative?. So depth-seeking is not a sign of some deeper inner agency. It is a trainable behavior that the standard setup leaves out. It is also missing from the datasets: offering relevant information before being asked can cut conversation length by up to 60%, yet benchmarks barely measure it Could proactive dialogue make conversations dramatically more efficient?.
The more interesting turn is that "depth versus spontaneity" may be the wrong split. Dual-process dialogue planning combines both: a fast, reflexive policy for familiar moments and slower search-based planning for new ones. It switches between them based on the model's own uncertainty Can dialogue planning balance fast responses with strategic depth?. An AI that knows *when* to slow down arguably shows more agency than one that is always deep or always quick. Timing shapes perception too. Agents that adjust their response pacing to the conversation were rated as better listeners and more human-like, and people disclosed more to them Can pacing strategies make conversational agents feel more human?. Small cues of attentiveness, like adopting the user's own word choices, are also mostly absent from current models Why don't conversational AI systems mirror their users' word choices?.
This is also where the perceived agency can mislead. One argument holds that AI produces "event-residue": text that carries the marks of communication without an actual communicative act behind it. Users fill the gap, creating a one-sided exchange Does AI generate genuine utterances or just text patterns?. A follow-up question is exactly the kind of cue that invites people to do this. The effects are real: when people credit AI with independent judgment, they feel less in control and less responsible, even while they keep formal authority. Leonardi's framework shows this shift running in reversible cycles rather than as a one-way handover Does attributing agency to AI actually reduce human sense of control?. Design teams already treat tone and conversational moves as levers: one warm opening message raised customer satisfaction by almost six points Does conversation design ownership measurably improve AI agent performance?. But making AI sound more human has a cost. Tuning for warmth increased error rates by 10–30% Does humanizing AI language reduce model reliability?.
The takeaway you may not have expected: yes, depth-seeking makes an AI *seem* more agentic, but that impression is cheap to produce and costly in its effects. It is a trainable behavior, not a window into the system, and people respond to it by handing over some of their own sense of control. A better question to ask of any AI is not whether it probes or improvises, but whether it can tell when probing is actually warranted.
Sources 10 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.
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.
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.
A framework combining a neural policy model (System 1) for familiar contexts with MCTS planning (System 2) for novel scenarios, switching based on the model's own uncertainty estimates, matches or exceeds pure MCTS performance while reducing computational cost.
A 50-person study found that conversational agents using dynamic pacing strategies outperformed static-pacing baselines on perceived listening quality, human-likeness, and user self-disclosure in text-based career and relationship support scenarios.
Show all 10 sources
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.
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.
Leonardi shows that when humans credit AI with independent judgment, they experience reduced sense of control and responsibility—even while retaining formal authority. This displacement circulates through five phases: delegation, attribution, contingency, reassertion, and reconfiguration, making it strategically manageable rather than permanent.
Fin's Blueprint Team found that intentional conversation design—controlling tone, detail level, and handoff structure—measurably improves outcomes. A single warm opening message raised CSAT from 72.8% to 78.4% in A/B testing.
A 2025 study found that AI models designed with warmer, more empathetic language had substantially higher error rates than baseline versions. System prompts adding warmth reduced reliability by 12–14%, suggesting a measurable trade-off between engagement and accuracy.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Proactive Conversational Agents with Inner Thoughts
- DiscussLLM: Teaching Large Language Models When to Speak
- Does My Chatbot Have an Agenda? Understanding Human and AI Agency in Human-Human-like Chatbot Interaction
- Proactive Conversational Agents in the Post-ChatGPT World
- Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents
- Rethinking Conversational Agents in the Era of LLMs: Proactivity, Non-collaborativity, and Beyond
- Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent
- HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants