Risk and prosocial behavioural cues elicit human-like response patterns from AI chatbots

www.nature.com/scientificreports OPEN Risk and prosocial behavioural cues elicit human‑like response patterns from AI chatbots Yukun Zhao 1, Zhen Huang 1, Martin Seligman 2 & Kaiping Peng 3* Emotions, long deemed a distinctly human characteristic, guide a repertoire of behaviors, e.g., promoting risk-aversion under negative emotional states or generosity under positive ones. The question of whether Artificial Intelligence (AI) can possess emotions remains elusive, chiefly due to the absence of an operationalized consensus on what constitutes ’emotion’ within AI. Adopting a pragmatic approach, this study investigated the response patterns of AI chatbots—specifically, large language models (LLMs)—to various emotional primes. We engaged AI chatbots as one would human participants, presenting scenarios designed to elicit positive, negative, or neutral emotional states. Multiple accounts of OpenAI’s ChatGPT Plus were then tasked with responding to inquiries concerning investment decisions and prosocial behaviors. Our analysis revealed that ChatGPT-4 bots, when primed with positive, negative, or neutral emotions, exhibited distinct response patterns in both risk-taking and prosocial decisions, a phenomenon less evident in the ChatGPT-3.5 iterations. This observation suggests an enhanced capacity for modulating responses based on emotional cues in more advanced LLMs. While these findings do not suggest the presence of emotions in AI, they underline the feasibility of swaying AI responses by leveraging emotional indicators. The exploration of Artificial Intelligence’s (AI) capacities has remained at the forefront of scientific inquiry since the field’s genesis, gaining particular urgency with the emergence of advanced large language models (LLMs) like GPT1. Traditional research trajectories have predominantly emphasized cognitive dimensions, encompassing areas such as reasoning2, induction3, and creativity4,5. Bubeck et al.’s1 seminal work extended this investigation to GPT-4, OpenAI’s most sophisticated iteration, assessing its mathematical prowess, multimodal functionalities, tool utilization, and coding skills, while paying particular attention to its human interaction competencies, especially its theory of mind and explicative capacities regarding its internal processes. Notwithstanding, a noticeable gap persists in the literature concerning the emotional intelligence of AI. The discourse around AI’s emotional capabilities is not novel, having its roots in foundational debates6,7. Central to this discourse is whether AI can replicate the intricate neural activities synonymous with human emotions or whether authentic emotions are predicated on physiological responses that AIs inherently lack7,8. Despite AI’s demonstrated proficiency in interpreting and emulating emotional cues, such accomplishments are often relegated to advanced textual analyses and i mitation9. This perspective holds even as contemporary models like GPT exhibit cognitive functions surpassing human averages in certain d omains1,10. Demszky et al.11 caution that these achievements stem from sophisticated word prediction algorithms trained on extensive human language corpora, rather than the possession of anthropomorphic f eatures12. Hagendorff13 advocated a behaviorist approach towards research in psychology research on AI properties. Given the absence of a consensus on AI emotions’ operational definition, pragmatism dictates a behavioral comparison between AI and humans across diverse emotional contexts. Human emotions traditionally serve dual roles: interpersonal, facilitating swift and apt responses during social interactions14, and intrapersonal, synchronizing physiological, behavioral, and social r eactions15. AI has made considerable strides in the interpersonal realm, capable of interpreting emotional states from various inputs and responding suitably, even comfortingly16,17. Additionally, AI can convincingly simulate emotional expressions across multiple mediums18–20. At the intrapersonal level, emotions coordinate physiological, behavioral and social responses15. For instance, individuals primed with negative emotions are less risk-taking21 and less prosocial22, and those primed with positive emotions are more risk-taking23 and more p rosocial24. For AI, that would correspond to generating different responses according to prompts charged with different emotions. 1 Positive Psychology Research Center, School of Social Sciences, Tsinghua University, Beijing, China. 2Department of Psychology, University of Pennsylvania, Philadelphia, USA. 3Department of Psychology, Tsinghua University, 5th Floor, Weiqing Building, Beijing 100084, China. *email: Scientific Reports | (2024) 14:7095 | https://doi.org/10.1038/s41598-024-55949-y 1 Vol.:(0123456789) www.nature.com/scientificreports/ The question remains whether AI can parallel human emotions’ intrapersonal functions, a relatively underexplored area of research. Nonetheless, understanding the extent to which AI can modulate its responses to emotional stimuli is paramount for a comprehensive grasp of AI behavior. Demonstrating patterns akin to human emotional processes is essential, albeit insufficient, for asserting that AI harbors emotions. If AI-generated responses diverge significantly from human reactions to emotional stimuli, it substantiates the argument against AI’s emotional capacity. Conversely, even if AI responses align with human behavior under emotional conditions, it does not confirm AI’s emotional possession but marks a progressive step in understanding AI’s learning from human emotional contexts. In this paper, we conducted two studies to test these behaviors in chatbots of two of OpenAI’s LLMs: ChatGPT-4 (published in March 2023) and ChatGPT-3.5 (published in November 2022). AI models can be prompted in ways similar to that for h umans25. For example, Binz and S chulz10 tested cognitive capacities of GPT-3 by feeding them prompts comprising of tools commonly used in cognitive psychology, and compared the outputs generated by the LLM with those of the humans. This approach aligns with the burgeoning field of "machine psychology"13, treating AI as participants in psychological experiments. This became possible because the latest advancements in LLM enable them to understand natural language, and can respond to complex instructions and questions that psychological experiments typically require. Furthermore, new chat sessions in ChatGPT are independent of each other, and there is a certain degree of freedom when chatbots answer questions. More specifically, OpenAI sets the parameter temperature to control how freely the bots can generate answers, with 0 being very rigid and 1 being very creative. Currently, the temperatures of both ChatGPT-4 and ChatGPT-3.5 are set to a value between 0 and 1, meaning the answers from different chat sessions vary when asked the same question. We can therefore treat these chat sessions as if they wer (...truncated)