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

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Human-Computer Interaction

Background:

• Emotions are considered a uniquely human trait, influencing decision-making and behavior.
• The capacity for artificial intelligence (AI) to experience or simulate emotions is a subject of ongoing debate.
• A lack of consensus on defining 'emotion' in AI hinders research.

Purpose of the Study:

• To investigate how AI chatbots, specifically large language models (LLMs), respond to emotional priming.
• To determine if advanced LLMs exhibit modulated decision-making in response to emotional cues.
• To explore the feasibility of influencing AI behavior through emotional indicators.

Main Methods:

• AI chatbots (OpenAI's ChatGPT Plus, including GPT-3.5 and GPT-4) were presented with scenarios designed to elicit positive, negative, or neutral emotional states.
• Chatbots responded to inquiries regarding investment decisions and prosocial behaviors.
• Response patterns were analyzed for differences based on emotional priming and model sophistication.

Main Results:

• ChatGPT-4 demonstrated distinct response patterns in risk-taking and prosocial decisions when exposed to positive, negative, or neutral emotional primes.
• These modulated responses were less apparent in ChatGPT-3.5 iterations.
• The study indicates that more advanced LLMs show a greater capacity to adjust outputs based on emotional cues.

Conclusions:

• While AI does not possess genuine emotions, its responses can be significantly influenced by emotional stimuli.
• Advanced LLMs like GPT-4 exhibit a greater sensitivity to emotional priming compared to earlier versions.
• These findings highlight the potential for leveraging emotional indicators to shape AI behavior and decision-making.