Conversational Health Interfaces in the Era of LLMs: Designing for Engagement, Privacy, and Wellbeing

Biases and Fairness in CUIs

One recurring theme was the bias embedded in foundation models and how it manifests in health conversations.

• Benevolent bias isn’t always benevolent: When agents framed responses positively (e.g., “You’ve achieved a lot at such a young age”), participants perceived them as patronizing or insincere. Forced positivity can drift into toxic positivity.
• Empathy vs. anthropomorphism: Overly empathetic or human-like responses sometimes triggered negative reactions (an “uncanny valley” effect). Context matters—transactional health advice differs from emotional support.
• Transparency matters: Users want to know how their history shapes responses. CUIs should clearly explain why certain suggestions are made.
• Safeguarding vs. misrepresentation: Attempts to make models “safer” can sometimes distort representation, introducing different perceived biases.

Open research questions

• How can we study bias and fairness ethically without harming participants?
• What triggers cause a chatbot to appear biased or unfair?
• How much does “personality” matter in user perceptions of fairness?

Stress Interventions and User Agency

Another group examined how CUIs might deliver stress interventions in office settings—a context where privacy and discretion are paramount.

Key insights

• Stress detection is tricky: Sensors may misinterpret energy as stress. Users should confirm or adjust predictions rather than being told “you are stressed.”
• Delivery must respect context: A desk worker in a meeting might want to defer an intervention. CUIs should ask, “Would you like to do this now or later?”
• Multi-modality is essential: Voice can be intrusive in shared spaces; offering text or silent modes increases accessibility.
• Customization at onboarding: Let users pre-set preferences (e.g., how to handle stress during meetings) and refine them post-intervention, rather than mid-stress.
• Transparency and trust: Clarify data sources, access control, and whether employers/colleagues can see information. Agency comes from informed choice.

Takeaway: Stress-support CUIs must empower users with control over when, how, and why interventions occur.

Proactivity in Exercise Support

The third group explored proactive CUIs for exercise, asking how interventions can be encouraging without being intrusive.

• A fine line between support and overbearing: Enthusiastic prompts may motivate some users but alienate others who feel patronized.
• Context changes everything: Illness, mood, location, and long-term goals shape what support is appropriate.
• Beyond metrics: Shift from rigid, goal-based prompts to reflective dialogues (e.g., “How are you feeling right now?”) to foster meaningful, long-term engagement.
• Ethical concerns: Proactive CUIs risk over-dependence or “optimization addiction.” How much intervention is too much?

This discussion underscored the need for proactive CUIs that listen, adapt, and evolve with users—balancing helpfulness with respect for autonomy.

Closing Thoughts

Across these sessions, one message was clear: Conversational Health Interfaces must walk a delicate line. They need to be proactive without being intrusive, empathetic without being condescending, and supportive without compromising user agency or privacy.

To get there, we must:

• Build transparency into every interaction
• Offer flexible modalities for diverse contexts
• Study bias with ethical, innovative methods
• Design with the lived experiences of users at the center

The workshop was only the beginning, but it underscored that the future of health CUIs depends not just on smarter models—but on smarter design choices that empower people to take charge of their health and wellbeing.