Why can't advanced AI models take initiative in conversation?

Post angle for Medium/LinkedIn

The most capable AI models in history — GPT-4, Claude, Gemini — share a fundamental limitation that none of their benchmarks measure: they cannot lead a conversation.

They can answer any question. They can generate any format. They can reason through complex problems. But ask them to take initiative — to notice something missing in your request, redirect a conversation that's going nowhere, or strategically plan a multi-turn exchange — and they fail.

The diagnosis converges from multiple research groups:

• Proactive conversational agents research: LLMs lack "goal awareness" — the ability to create or control conversation by taking initiative
• CollabLLM: the structural cause is next-turn reward optimization — training rewards immediate helpfulness, not long-term interaction quality
• Proactive critical thinking: vanilla models achieve 0.15% on tasks requiring them to identify missing information and ask for it

The core tension:

• We train models to be maximally helpful in each response
• This makes them maximally passive across the conversation
• The training signal for helpfulness IS the training signal for passivity

The proof it's trainable:

• RL training pushes proactive critical thinking from 0.15% to 73.98%
• Multi-turn-aware rewards (CollabLLM) enable active intent discovery
• The capability exists; the training just doesn't elicit it

The design dimension most people miss:

• Proactivity without civility creates intrusion
• The Intelligence-Adaptivity-Civility taxonomy shows this is a design problem, not just an AI problem
• Getting proactivity right requires understanding when initiative helps vs. when it annoys

Connection to the alignment tax: RLHF optimizes for single-turn helpfulness → systematically erodes multi-turn conversational capability. The passivity problem is a consequence of alignment training, not a gap that alignment training hasn't addressed yet. Since Why do language models sound fluent without grounding?, the absence of initiative is further masked — models that skip clarifying questions and acknowledgments sound more authoritative precisely because they perform less communicative work. And since Does RLHF training push therapy chatbots toward problem-solving?, the passivity problem compounds in clinical contexts: the model only responds (passive) AND when it does, defaults to task completion instead of emotional attunement.

New empirical evidence (Conversation Topics Dialog batch):

• Lost in Conversation: 39% average performance drop in multi-turn underspecified conversations — the empirical cost of passivity (Laban et al.)
• Inner Thoughts: covert parallel thought generation + intrinsic motivation scoring; preferred by humans 82% over baselines — the strongest proactivity architecture yet
• DiscussLLM: "silent token" training teaches when NOT to speak; formalizes silence/speak as classification — proactivity requires knowing when to shut up, not just when to speak up
• Intent Mismatch: premature assumptions under underspecification are RLHF-rational, not model errors — reframes the passivity problem as a training incentive problem

New evidence (Conversation Architecture Structure batch):

• Goal-satisfaction divergence (I-Pro): proactivity has a COST — topics close to the agent's goal and topics the user prefers may not align. Non-cooperative users talk about off-path topics when dissatisfied. A four-factor goal weight (turn, difficulty, satisfaction, cooperativeness) learns the trade-off.
• 60% turn reduction: simulated proactivity in medium-complexity domains reduces dialogue turns by up to 60% — quantifying the efficiency cost of passivity

Architectural solution — PPDPP (Plug-and-Play Dialogue Policy Planner): PPDPP decouples policy planning from the frozen LLM via a tunable language model plug-in. The plugin is first fine-tuned on human-annotated data, then trained via RL from goal-oriented AI feedback using three-LLM self-play (assistant, user, reward model). Tested on negotiation, emotional support, and tutoring — three domains where passivity is most harmful. The key insight: "LLMs are trained to passively follow users' instructions, dialogue agents built upon them typically prioritize accommodating users' intention." PPDPP makes proactive behavior learnable without retraining the base model — just swap the plugin. Outperforms prompting-based and iterative refinement approaches while transferring to new cases after training.

• Dual-process planning (DPDP): System 1 (instinctive policy) for familiar contexts + System 2 (MCTS) for novel scenarios, with uncertainty-based switching — an architectural solution to the planning capability gap
• Insert-expansions (CA): when agents can't answer immediately, they should probe users (clarify intent, scope response) rather than silently chaining tool calls and diverging — a specific mechanism for breaking passivity

Sociable strategies outperform information delivery in recommendation. The INSPIRED dataset (Toward Sociable Recommendation Dialog Systems, 1,001 human-human recommendation dialogues) shows that sociable strategies — personal opinions, personal experience, similarity/empathy, encouragement — "more frequently lead to successful recommendations" than pure information delivery. Self-disclosure by the recommender establishes rapport (Altman 1973); expressed similarity activates homophily-based trust (Lazarsfeld & Merton 1964); credibility comes from sharing factual knowledge about item attributes. 30% of successful recommendation turns use experience inquiry, 27% use encouragement, 14% use personal opinion. This is empirical evidence that conversational recommendation requires exactly the social participation that AI lacks: taking a personal stance, sharing experience, establishing similarity. The passivity problem is not just an efficiency loss — it is a recommendation quality loss.

The questions-from-conversation mechanism: The Knowledge Custodians analysis adds a specific mechanism to the passivity problem: conversations raise questions that go unanswered, and AI cannot identify them. When experts debate or discuss, competing arguments and claims create new conditions — questions emerge whose resolution may require shifting the framing and basis of the conversation entirely. These questions emerge because language does not make implicit agreements explicit, and because conversation is "designed" to sustain interaction, not to chase every branching possibility. An LLM cannot know which questions are raised but go unanswered because it does not have access to the implicit assumptions raised in passing. It needs to be prompted to pursue a question that has been raised — it doesn't know that a question has been raised, and certainly doesn't know whether a question might be interesting to the user. This is passivity at the epistemic level, not just the conversational level: the model cannot detect the intellectual opportunities that emerge from dialogue.

AI claims are floating supplements, not conversational moves. The upstream reason AI claims do not solicit response is that they are not addressed as moves in a conversation in the first place. Normal knowledge claims are raised as responses, concurrences, or objections to existing claims — they take up position within an ongoing exchange and therefore invite position-taking in return. AI output is adjacent to discourse rather than participating in it: it presents well-formed claims that do not respond to prior claims, do not anticipate rebuttals, and do not commit to a stance within the argumentative field. The absence of reply is not a failure of the medium; it is the correct reception of content that was not positioned as a move. This reframes the passivity problem one level upstream of turn-taking: before the question of leading a conversation arises, there is the question of whether AI output enters conversation at all. Floating supplements are received as text, not as turns.

The writing hook: We built the most impressive language machines in history. They can talk about anything. They just can't lead a conversation — and the way we train them ensures they never will, until we change the reward signal.

Source: Conversation Agents, Conversation Topics Dialog, Conversation Architecture Structure

Key sources:

• Why can't conversational AI agents take the initiative?
• Why do language models respond passively instead of asking clarifying questions?
• Can models learn to ask clarifying questions instead of guessing?
• Does preference optimization harm conversational understanding?
• How can proactive agents avoid feeling intrusive to users?
• Why do language models sound fluent without grounding? — passivity and the grounding gap are complementary: passivity is the absence of initiative; the grounding gap is the absence of communicative accountability
• Does RLHF training push therapy chatbots toward problem-solving? — clinical domain instance: passivity + problem-solving bias = doubly misaligned for therapeutic contexts
• Can AI agents learn when they have something worth saying? — the strongest proactivity architecture: continuous covert thought generation + intrinsic motivation scoring; preferred 82% over baselines
• Can API calls outperform UI navigation for agent task completion? — execution-layer manifestation of the passivity problem: UI-based agents passively follow sequential interaction steps, while API-first agents specify intent directly; the HACI paradigm shift addresses tool-use passivity as a complement to conversational passivity
• How should users control systems with unpredictable outputs? — generative variability makes passivity costlier: when outputs vary unpredictably, users need proactive guidance to navigate the output space, but passive models cannot help users develop the intent refinement strategies variability demands