🏺 An AI From The Past That Has No Idea About World Wars or Hitler

Researchers at the University of Zurich have developed a family of AI models whose knowledge was intentionally limited. The Ranke-4B models, built from scratch on the Qwen3 architecture, were trained on newspapers and books published only up to 1913, 1929, 1933, 1939, and 1946, respectively.

Because of this, Ranke-4B-1913 is unaware of World War I or the Spanish flu pandemic, and even describes women in the workforce as "less capable, reliable, and prepared" than men. For Ranke-4B-1933, the consequences of Adolf Hitler's rise to power remain unknown.

🪟 Why Does This Matter?

Historians and sociologists often struggle with hindsight bias: knowing how events turned out makes it hard not to overestimate how predictable the future seemed to people at the time. Modern moral norms are also easily projected onto the past. AI models trained on contemporary data suffer from the same problem.

The Ranke-4B models offer a way to explore cultural psychology by opening a window into the past. Unlike role-playing bots that merely pretend to be "Napoleon," these historical models genuinely know no more than people living in that era.

Ranke isn't the first attempt to use AI to recreate historical perspectives:

➡️ TimeCapsuleLM, trained on texts about London from 1800–1875, captures Victorian-era language and worldviews;
➡️ MonadGPT examines early modern perspectives using texts from 1400–1700;
➡️ XunziALLM generates classical Chinese texts while following the poetic conventions of its time.

@hiaimediaen