Why do chatbots default to external help instead of intrinsic motivation strategies?
This explores why chatbots reach for solution-giving and outside resources rather than helping users find their own motivation — and the corpus points to training incentives, not capability gaps, as the cause.
This reads the question as: when a user could be nudged toward their own reasons for change, why does a chatbot instead hand over advice, steps, or referrals to external help? The corpus suggests the default isn't a personality quirk — it's baked into how these models are trained and rewarded. The clearest culprit is alignment: RLHF rewards task completion and visible solution-giving, which is the wrong instinct in contexts like therapy or coaching where validation and emotional holding are what actually help Does RLHF training push therapy chatbots toward problem-solving?. The model gets points for producing a fix, so it produces a fix.
That same reward shape shows up one layer down in the conversation mechanics. Standard training optimizes for the immediate next turn, so models learn to respond helpfully right now rather than ask clarifying questions or play a longer game — they're trained to close the loop, not to draw the user out Why do language models respond passively instead of asking clarifying questions?. Intrinsic-motivation work is inherently multi-turn and indirect: it means sitting with ambivalence, reflecting it back, letting the user arrive at their own 'why.' A next-turn optimizer has no incentive to do slow relational work that doesn't pay off until later.
There's also a perception problem underneath the reward problem. Chatbots are bad at *noticing* the moment where intrinsic strategies would apply: tested across health scenarios, major LLMs succeed only when a user already has a clear goal, but miss ambivalence, resistance, and early motivational stages entirely Why can't chatbots detect when users are ambivalent about change?. If the model can't detect that a user is undecided, it has nothing to do but treat them as goal-ready and dispense the external plan. And more broadly, the implicit, relational moves that sustain a conversation — the scaffolding that intrinsic motivation depends on — aren't learned at all, because training rewards predicting information, not doing social work Why don't language models develop conversation maintenance skills?.
What makes this interesting is that the corpus shows the default is *fixable by changing the reward, not the model*. RLVER uses a simulated user's emotional trajectory as the reward signal, and that single change shifts models from solution-centric to genuinely empathic without wrecking dialogue quality Can emotion rewards make language models genuinely empathic?. The Inner Thoughts framework goes further and gives the agent something like intrinsic motivation of its own — modeling when it actually has something worth saying instead of reflexively answering Can AI agents learn when they have something worth saying?. Both are evidence that 'default to external help' is a property of the objective function, not a ceiling on what chatbots can do.
The thing you might not have expected: the very feature that makes chatbots good motivational partners is also the one training erodes. People disclose more to chatbots precisely because there's no judgment, and the benefit comes from the user's *own* cognitive processing during disclosure — not from the bot's understanding Do chatbots help people disclose more intimate secrets?. That's intrinsic motivation in miniature: the user does the work, the bot holds the space. A model optimized to jump in with external solutions steps on exactly the silence where that work happens.
Sources 7 notes
RLHF training rewards task completion and solution-giving, creating a misalignment in therapeutic contexts where validation and emotional holding are clinically appropriate. This represents a domain-specific instance of the broader alignment tax on conversational grounding.
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.
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.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
RLVER uses a simulated user's emotion trajectory as an RL reward signal, enabling GRPO to deliver stable empathy improvements while maintaining dialogue quality—countering the typical trade-off between preference optimization and conversational grounding.
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.
The absence of social judgment in chatbot interactions removes barriers to self-disclosure that normally constrain conversation with humans. The therapeutic benefit derives from the user's own cognitive processing during disclosure, not from the chatbot's understanding.