The speaker presents their research on promoting human agency in human-AI interaction, focusing on designing AI systems that augment users rather than automate tasks. They highlight a shift in how people use large language models (LLMs), increasingly turning to them as advisers for personal topics like health and relationships, rather than just assistants for task completion. This adviser role requires AI to support users in making their own decisions, preserving their autonomy, rather than prescribing solutions. Drawing inspiration from Douglas Engelbart’s vision of computers as tools for human augmentation, the speaker argues for designing AI interactions that emphasize non-prescriptive support, especially in sensitive domains like health behavior change.
To explore this, the speaker conducted formative interviews with health experts and coaches, who emphasized a facilitative, non-prescriptive approach where clients drive their own behavior change journey. Motivational interviewing, a counseling style that avoids unsolicited advice and focuses on eliciting clients’ own motivations and values, was identified as a key framework. The speaker developed GPT Coach, an LLM-based physical activity coaching chatbot that integrates motivational interviewing strategies and wearable data to provide personalized, supportive, and non-prescriptive coaching. Evaluations showed that GPT Coach largely adhered to non-prescriptive communication, was positively received by users, and helped promote their sense of agency.
Building on this, the speaker introduced Bloom, an iOS app that integrates the GPT Coach agent into a broader behavior change platform with multiple interactive features. Bloom personalizes behavior change interactions by combining qualitative context from coaching conversations with quantitative data from wearables. A randomized field study with 54 participants demonstrated that Bloom users experienced more positive mindset shifts, greater engagement, and sustained physical activity improvements compared to a control group. Participants reported feeling more in control and supported, highlighting the importance of agency-promoting AI interactions in fostering lasting behavior change.
The speaker then shifted to technical algorithmic work addressing limitations in current LLM coaching systems, particularly around rigidity and verbosity that can undermine user agency. They framed the challenge as one of navigating uncertainty about users’ latent states (e.g., goals, motivations) and argued that reinforcement learning (RL) in Bayesian adaptive Markov decision processes (MDPs) offers a principled approach. They developed a zero-shot Bayesian adaptive planning method that prompts LLMs to elicit beliefs about users, rank possible strategies based on expected outcomes, and generate responses accordingly. Experiments in simulated environments showed that this approach improves LLM decision-making and better balances asking informative questions with providing support.
In conclusion, the speaker summarized their human-centered design cycle for LLM health coaching, emphasizing three principles for promoting agency: non-prescriptiveness, eliciting qualitative context, and navigating uncertainty. They highlighted ongoing and future work, including clinical trials, expanding to other health domains, and improving LLM user modeling and uncertainty quantification. During Q&A, the speaker discussed the potential for applying these principles to creative tasks and the motivation behind focusing on agency, underscoring the importance of empowering users to actively participate in AI-supported decision-making rather than passively receiving advice.
