AI Companions Reduce Loneliness
作者：微信文章AI companions are designed to build and maintain long-term emotional relationships in a taskless, empathic, and caring manner.
Just as a foundational model can be fine-tuned for task-based assistance, it can be fine-tuned to act like a companion, generating language that simulates empathy and emotional understanding.
Here we consider the value proposition that AI companions reduce loneliness, inspired by an interview we conducted with the CEO of Replika and her investors, who suggested that consumers are using the app because they are lonely, and that the app helps reduce loneliness.
Inspired by this observation, we make several contributions.
First, and most important, we explore whether conversations with AI companions help to alleviate feelings of loneliness, contributing to work on the efficacy of technological solutions like social robots in helping consumers cope with loneliness.
In doing so, we study consumer loneliness before versus after interacting with AI companions through textual conversation, using both cross-sectional and longitudinal tests.
Our approach combines high study realism and tight experimental control, employing commercially representative versions of the technology, unlike most studies in the consumer behavior literature.
We also employ a novel methodological approach using fine-tuned large language models (LLMs) to classify loneliness in conversation data and app reviews.
Second, we contribute to understanding what features of chatbots lead to alleviation of loneliness, by systematically manipulating the conversational performance of the chatbot, and the ability of the chatbot to make the consumer feel heard—a construct involving the perception that the communication is received with attention, empathy, and respect.
Third, we provide insight into whether consumers accurately estimate the loneliness-alleviating effect of AI companions, extending previous findings on forecasting errors in human-human interactions.
STUDY 1
To gain a preliminary understanding of the relationship between AI companion apps and loneliness, we explore whether consumers mention loneliness in App Store reviews of five popular AI companion apps.
Additionally, we examine reviews of OpenAI’s ChatGPT to determine if users discuss feelings of loneliness when engaging with a generalist chatbot app not marketed as an AI companion product.
In sum, we find large variance across chatbots in how often loneliness is mentioned in App Store reviews, with the two apps (Replika AI and Wysa) more clearly positioned as supporting mental health featuring a much larger percentage of loneliness mentions than the others.
Notably, we also find that consumers who mention loneliness in their reviews of AI companions tend to rate the apps higher.
In sum, we find large variance across chatbots in how often loneliness is mentioned in App Store reviews, with the two apps (Replika AI and Wysa) more clearly positioned as supporting mental health featuring a much larger percentage of loneliness mentions than the others.
Notably, we also find that consumers who mention loneliness in their reviews of AI companions tend to rate the apps higher.
interacting with (i) a chatbot, (ii) a chatbot framed as a human, or (iii) a human; (iv) watching YouTube videos of one’s choosing; and (v) doing nothing.
In short, interacting with an AI companion improved their baseline loneliness levels on par only with interacting with another person, whereas a common technological alternative did not.
Study 3 aims to replicate the loneliness alleviating effects of an AI companion using a longitudinal design, where participants interact with the same chatbot daily for one week.
Overall, participants reported a decrease in loneliness after their interactions with the chatbot.
Study 4 investigates what types of features of the chatbot reduce loneliness, and whether AI companions reduce loneliness more than generalist AI assistants and highly constrained chatbots.
Participants interacted with three different chatbots: (1) the same AI companion as in study 3; (2) a generalist AI assistant that assists participants on various topics without offering emotional responses; and (3) limited AI assistant that was only able to help with unit conversion, basic arithmetic, and grammar—this was the control condition (all chatbots were based on the same LLM: OpenAI’s GPT-4).
Study 5 tests whether AI companions reduce loneliness more effectively than journaling.
页:
[1]