How Personas Can Influence Agents to Play Split or Steal

- This paper studies how AI agents behave in the classic "Split or Steal" game (also known as Prisoner's Dilemma), where two players each choose to either cooperate and split a prize or defect and try to steal it.<br>- The researchers gave large language models (LLMs) like GPT-4 different "personas" by prompting them with background traits (e.g., "you are a selfish trader") before playing the game. They found that persona descriptions strongly swayed whether the AI chose to cooperate or defect.<br>- This matters because LLMs are increasingly used as autonomous agents in settings like negotiations, auctions, and online marketplaces where strategic decision-making is required. If simple persona prompts can reliably change their behavior, it raises questions about manipulation, fairness, and reliability.<br>- The paper builds on a growing body of research showing that LLMs are not purely "rational" — they are influenced by framing, roleplay, and conversational cues in ways that mirror human biases.