Researchers at the Max Planck Institute and partner institutions have developed a method called "Time Machine Experiments" that trains AI models on historical text corpora with hard cutoff dates—so a model might be trained only on materials available through 1920, another through 1930, and so on. The core insight is elegant: this lets you ask questions of AI systems that genuinely don't have access to information from later periods, creating something closer to an actual historical perspective. Traditional approaches hit dead ends—archival records can't answer new questions you pose to them, and actual people from the past are now shaped by everything that happened afterward, making their recollections unreliable guides to what people actually thought at the time. A historically bounded AI, by contrast, can engage with novel questions using only the knowledge available in its training period. The method opens experimental possibilities that weren't there before. You can ask a 1920-bounded model questions that didn't exist in 1920, or probe its reasoning about events it has no information about. The researchers tested whether people today could meaningfully interact with these historical perspectives. They had participants ask questions and examine how the models responded, measuring effects on beliefs and behaviors. This transforms the past from something you read about into something you can actually interrogate—you get responses that reflect genuine historical ignorance rather than hindsight bias. The practical payoff matters for understanding both history and human cognition. You can test how historical beliefs would have responded to different information, or study how people today react when confronted with authentic historical uncertainty. The approach is justifiable because it prevents the model from doing worse than history itself—a 1920-bounded model can't accidentally reference 1945, which a standard model might do. The researchers grounded this in work from the Max Planck Institute on human development and decision-making, pulling in expertise from economics, psychology, and AI ethics.