AI agents learn to see the same scene differently based on what happened before

Researchers built a minimal artificial agent that changes how it perceives identical observations depending on its history — what it has experienced shifts what it sees, not just how it acts. This suggests AI systems can develop something like perspective or stance, where past events leave measurable traces in how the agent's internal model interprets the present.

Most AI agents treat each observation as a fresh input — the same image means the same thing every time. This paper shows a simpler mechanism: if you let an agent's internal state feed back into its perceptual layer and update slowly based on experience, it develops what looks like a stable viewpoint that colors how it interprets the world. The finding is structural, not behavioral. The agent's actions stayed consistent, but its internal representation of identical inputs diverged based on history. This matters because it identifies a minimal way for artificial systems to build persistent perspective without needing complex external memory or attention mechanisms.