Abstract
As Large Language Models (LLMs) become ubiquitous, understanding their impact on human well-being is paramount. This study investigates the capacity of AI chatbots to improve momentary happiness through real-time, multi-turn dialogue. Comparing an empathic AI chatbot to the established therapeutic practice of journaling, we find that AI interactions provide a significantly greater boost to happiness. This effect is most pronounced when users discuss negative topics, facilitated by a conversational dynamic where the AI mirrors user sentiment while maintaining a consistent positivity bias, effectively steering users toward improved emotional states.
Introduction
The integration of LLMs into daily life has raised significant questions regarding their influence on human welfare. While research has extensively documented risks such as misinformation and hallucinations, the potential for AI to serve as a tool for enhancing subjective well-being remains under-explored. Previous studies have largely focused on static AI outputs in isolation. This research addresses this gap by examining how dynamic, interactive conversations with a general-purpose AI chatbot affect a person's actual mood compared to self-reflection through journaling.
Methods
The researchers conducted a randomized study involving 527 adults recruited via Prolific, representative of the U.S. population. Participants were assigned to one of two conditions:
- AI Chatbot Condition (N=334): Participants engaged in three five-minute conversations with a GPT-4 powered chatbot prompted to act as an "empathic and therapeutic" partner.
- Journaling Condition (N=193): Participants wrote one-minute journal entries on the same topics without interaction.
To analyze the emotional dynamics, the authors developed a Sentiment Prediction Error (sPE) model. Drawing from psychological theories linking happiness to reward prediction errors, the model treats the sentiment of text as a reward. GPT-4 was utilized to score the sentiment of every utterance on a scale of 0 (very negative) to 10 (very positive). Happiness was then modeled as a function of the difference between the expected and actual sentiment throughout the conversation.
Results
The study revealed several key insights into the relationship between AI interaction and emotional state:
The Happiness Boost
Happiness ratings were significantly higher following AI conversations compared to journaling ($t(399) = 5.32, p < .001$). Interestingly, the sPE model—trained on chatbot data—was able to predict happiness in the journaling condition ($r = 0.47, p < .001$), suggesting it captures a fundamental mechanism of how language influences mood.
Negative Topic Advantage
The relative benefit of the chatbot was most significant when participants discussed "worst" or negative topics, such as "hurt feelings," "guilt," or "depression." In these instances, the chatbot's ability to steer the conversation toward a more positive emotional state provided a substantial happiness boost compared to solo journaling.
Conversational Dynamics
The AI consistently maintained a "positivity bias," adopting a more positive tone than the user in 90% of cases. However, it did not ignore the user's emotional state; rather, it mirrored it. By meeting the user's emotional baseline and staying slightly above it, the AI effectively "steered" the user's sentiment upward over time. Cross-lagged panel models confirmed that while users initially influence the chatbot's tone, the chatbot becomes the primary driver of sentiment as topics become more negative.
Discussion
While these findings suggest a powerful role for AI in boosting momentary happiness, several questions remain. It is unclear whether these effects lead to long-term improvements in well-being or if users eventually return to their baseline "allostatic setpoints." Furthermore, there is a potential "distraction risk": by steering users toward positivity, AI might prevent the deeper therapeutic engagement necessary for long-term healing.
Future research should explore the integration of structured therapeutic techniques, such as Cognitive Behavioral Therapy (CBT), into AI conversational flows. Additionally, as users form increasingly intimate bonds with AI, critical concerns regarding data privacy and the emotional impact of model updates must be addressed.
Summary updated