Can what an AI “sees” change how it behaves? That's the fascinating question explored by new research examining how visual personas impact the behavior of large language models (LLMs). Researchers gave LLMs fictional avatar images as visual personas, and then observed their negotiation strategies in a simulated game. The results revealed that LLMs with more aggressive-looking avatars tended to make bolder, more self-serving offers in the game. Interestingly, these AI agents also seemed to assess the aggressiveness of their opponents' avatars, adjusting their tactics accordingly. When faced with a less aggressive-looking opponent, they took a more dominant stance. Conversely, they were more likely to concede when facing opponents with intimidating avatars. This suggests that, similar to humans, visual cues play a significant role in how LLMs perceive social dynamics and make decisions. The implications are far-reaching. As LLMs become more integrated into our lives, understanding how visual information shapes their interactions is crucial for responsible development. One model, GPT-4o, demonstrated a greater ability to adapt its behavior based on the perceived aggressiveness of its own and its opponent's avatar, exhibiting a sort of “kiss up, kick down” dynamic. This raises important ethical questions: if AI can be influenced by visual cues, how do we prevent biased or harmful behaviors from emerging? While this research focused on aggressiveness, it opens doors to explore how a broader range of visual traits could shape LLM behavior. Imagine AIs trained with personas that reflect empathy, helpfulness, or even humor—could this lead to more human-like and positive interactions? There’s much more to uncover, but this research provides a compelling glimpse into the complex interplay between visual perception and AI behavior, highlighting both the potential and the challenges that lie ahead.