Designing Trustworthy LLM-based Wellbeing Recommendation through Controllable Interaction

The Problem with Implicit Adaptation

Large Language Models (LLMs) are increasingly used to provide personalized guidance for physical activity, mental health, and stress management. While these models offer fluent, context-aware interaction, their behavior is typically shaped by implicit mechanisms like prompting and fine-tuning. The authors argue that this approach is problematic for wellbeing applications, where recommendations directly influence a user's health and long-term outcomes. Because the interaction style emerges from the model's generative process rather than explicit design, it is difficult to inspect, control, or ensure that the system is acting in the user's best interest.

A Framework for Controllable Interaction

To address these issues, the authors propose shifting from viewing recommendation as a prediction problem to viewing it as an interaction design problem. They introduce a modular architecture that separates the generative model from an explicit 'interaction policy layer.' This layer enforces configurable constraints that govern how the system communicates with the user. By structuring conversational behavior through these parameters, designers can move away from opaque, emergent behavior toward a system that is transparent and intentionally aligned with human wellbeing.

Key Interaction Constraints

The paper identifies four primary constraints that should be explicitly managed in wellbeing-oriented LLM systems:

• Guidance Strategy: Determining whether the system positions itself as a directive authority or a collaborative partner.
• Degree of Directness: Adjusting the assertiveness of language (e.g., 'you should' vs. 'you might consider') to balance persuasion with user autonomy.
• Explanation Style: Defining how the system provides reasoning, ensuring that explanations are grounded in data rather than just being fluent but potentially misleading.
• Responsibility Framing and User Control: Establishing clear boundaries for decision-making authority, including mechanisms for users to override suggestions or adjust their goals.