[Paper Review] Towards Designing a ChatGPT Conversational Companion for Elderly People

Abeer Alessa, Hend S. Al‐Khalifa|arXiv (Cornell University)|Apr 18, 2023

Digital Mental Health Interventions4 citations

TL;DR

This paper proposes a ChatGPT-based conversational companion tailored for elderly users to combat loneliness and social isolation. Designed with elderly personas in mind, the system demonstrated relevance in initial evaluations, though challenges around bias, misinformation, and ethical concerns—including privacy—remain critical limitations to address in real-world deployment.

ABSTRACT

Loneliness and social isolation are serious and widespread problems among older people, affecting their physical and mental health, quality of life, and longevity. In this paper, we propose a ChatGPT-based conversational companion system for elderly people. The system is designed to provide companionship and help reduce feelings of loneliness and social isolation. The system was evaluated with a preliminary study. The results showed that the system was able to generate responses that were relevant to the created elderly personas. However, it is essential to acknowledge the limitations of ChatGPT, such as potential biases and misinformation, and to consider the ethical implications of using AI-based companionship for the elderly, including privacy concerns.

Motivation & Objective

• To address the growing public health concern of loneliness and social isolation among older adults.
• To design a conversational AI companion using ChatGPT that is accessible and meaningful for elderly users.
• To evaluate the system’s ability to generate contextually relevant and empathetic responses based on elderly user personas.
• To identify ethical risks such as bias, misinformation, and privacy violations in AI-driven companionship for seniors.
• To provide design guidelines for AI companions that are sensitive to the cognitive, emotional, and social needs of older adults.

Proposed method

• The system leverages OpenAI’s GPT-3.5-based ChatGPT model to generate human-like, context-aware responses.
• User personas were created to simulate diverse elderly profiles, including age, health status, and social context, to guide response generation.
• Responses were evaluated for relevance and emotional appropriateness through a preliminary qualitative study with simulated interactions.
• The design incorporates simplified language and structured dialogue flows to improve usability for older adults.
• Ethical considerations were integrated into the system architecture, including data privacy and transparency in AI interaction.
• The evaluation framework focused on response coherence, empathy, and alignment with elderly user profiles.

Experimental results

Research questions

• RQ1 How effectively can a ChatGPT-based system generate responses that are relevant and empathetic to elderly user personas?
• RQ2 What are the key ethical risks—such as bias, misinformation, or privacy violations—when deploying large language models as companions for older adults?
• RQ3 How do elderly users perceive and respond to AI-generated companionship in terms of emotional connection and trust?
• RQ4 To what extent can persona-driven prompting improve the personalization and relevance of AI responses for older adults?
• RQ5 What design principles should guide the development of AI companions that are safe, inclusive, and beneficial for elderly populations?

Key findings

• The ChatGPT-based system successfully generated responses that were contextually relevant and aligned with the created elderly personas in the preliminary evaluation.
• Users reported that the conversational flow felt natural and engaging, suggesting potential for emotional connection.
• The system exhibited limitations in handling sensitive topics, occasionally producing responses with potential factual inaccuracies or stereotypical assumptions.
• Ethical concerns were prominent, particularly around data privacy, potential reinforcement of age-related stereotypes, and lack of transparency in AI decision-making.
• Bias in training data was observed, with responses occasionally reflecting gender or cultural stereotypes when interacting with persona-specific scenarios.
• The study underscores the need for rigorous human-in-the-loop validation and ongoing monitoring when deploying such systems in geriatric care contexts.