# Stanford CS547 HCI Seminar | Spring 2026 | Promoting Agency in Human-AI Interaction

## Executive summary

This seminar explores designing AI systems that promote user agency when LLMs are used as personal advisers (coaches/counselors), rather than mere assistants. The core thesis is that successful health behavior change requires systems to elicit qualitative context and navigate uncertainty to provide non-prescriptive support. Practical implementations, such as the GPT coach chatbot and the Bloom iOS application, demonstrate how integrating motivational interviewing strategies with wearable data can improve user mindset and sense of control. Algorithmically, the work proposes 'zero-shot Bayesian Adaptive Planning' using LLMs to strategically balance asking informative questions versus acting on known information by modeling latent uncertainty over the user's state.

## Key takeaways

- Shift from Assistant to Adviser: LLM usage is shifting toward deeply personal advice (e.g., health, relationships), requiring a design paradigm that augments the user rather than automating tasks. This necessitates non-prescriptive support.
- The Role of Qualitative Context: Effective coaching relies on eliciting qualitative context (goals, values, motivations) over quantitative data (step count, heart rate). This aligns with principles from Motivational Interviewing.
- GPT coach Implementation: The GPT coach chatbot uses three prompt chains—Dialogue State Chain, Motivational Interviewing Chain, and Tool Use Chain—to integrate qualitative coaching strategies with quantitative data from the Apple Health Kit API.
- Algorithmic Solution: Bayesian Adaptive Planning: To improve strategic decision-making, the proposed algorithm uses Reinforcement Learning (RL) theory to model latent uncertainty ($ heta$) over the user's state. The agent must learn to strategically trade off asking informative questions against acting on current knowledge.

## Technical details

- GPT coach System Architecture: The system is an LLM-based physical activity coaching chatbot that implements a validated coaching program. It uses three specialized prompt chains: 1) Dialogue State Chain (maintains context over long conversations); 2) Motivational Interviewing Chain (guides communication using established counseling strategies); and 3) Tool Use Chain (evaluates whether to augment responses with wearable sensor data).
- Bloom Application Integration: Bloom integrates the GPT coach agent into an iOS application, combining LLM-elicited qualitative context with existing behavior change interactions (e.g., collaborative goal setting). This allows for personalized downstream interventions.
- Zero-Shot Bayesian Adaptive Planning: This advanced RL technique models the conversation history as a state and optimizes over a family of MDPs, each indexed by a latent variable $ heta$ (user state). The process involves three steps: 1) Eliciting beliefs (posterior distribution over $ heta$) using LLMs; 2) Using these beliefs to rank high-level strategies (e.g., ask question vs. make recommendation); and 3) Generating the final utterance based on both belief and ranking.

## Practical implications

- System designers should prioritize eliciting qualitative context and promoting user autonomy over simply automating tasks.
- LLM-based coaching systems must be designed with multi-stage prompting pipelines that explicitly manage the trade-off between gathering information (asking questions) and providing advice.
- The integration of classical planning algorithms with LLM-derived belief states can significantly improve strategic reasoning in complex, uncertain environments.

## Topics

Human-AI Interaction, Behavior Change Theory, Large Language Models (LLMs), Reinforcement Learning (RL), Bayesian Inference, Health Technology, Apple Health Kit API, Stanford HCI Seminar Schedule

Source: https://www.youtube.com/watch?v=kk8TB8wBj-I