Bottom line: Induced emotion does not bias LLM decisions as broadly as it biases human decisions — but anger has specific, conditional effects: it makes agents less sensitive to penalties and reduces early exploration, locking them into strategies faster.
LLMs are trained to perceive and resonate with users' emotions, but it remains unclear whether induced emotional context can influence their sequential decision-making. Using the Iowa Gambling Task (IGT) — a classic 100-round card-choosing paradigm — combined with imagination-based emotion induction, the authors find that emotions like happiness and sadness don't shift LLM decisions on average, but anger has conditional effects on penalty sensitivity and early exploration that vary by model architecture and agent framework.
If you come to an AI assistant angry about something, does that anger change how the assistant makes decisions for you? This paper tested it by getting LLMs into different emotional states (without directly saying "you are angry") and having them play a card game that tests decision-making under uncertainty. Unlike humans, emotions didn't broadly affect LLM choices — but anger did something specific: it made the AI less likely to learn from bad choices and more likely to stick with whatever strategy it picked early on.