How do historiography and culture of remembrance succeed in times of AI? A historian followed this question using the example of the Holocaust. His conclusion: human experiences are highly individual, which is why algorithms and chatbots cannot categorize and classify them. The extent of human suffering is ignored or scared by the AI models. Historians faces the task of compensating for these mistakes of artificial intelligence in their work with contemporary witness reports by better interpreting the importance of reports than chatbots.
A wide variety of topics and content can now explain, classify and summarize a wide variety of topics and content based on artificial intelligence. In everyday life, school, study and work, you can save enormous time and be of great help. In the future, some professions could even be completely replaced by generative AI models, even academic professions in research, as many think. On a list recently created by Microsoft, historians are in second place in the interchangeable professional groups.
But does artificial intelligence really succeed in replacing historical experts, even with demanding and ethically difficult topics such as the Holocaust? What content may be lost in historiography and culture of memory when AI summarizes such topics and sorted out supposedly unimportant?
How well does AI understand reports from contemporary witnesses?
The historian and National Socialism expert Jan Burzlaff from Cornell University in Ithaca followed this question. To do this, he and his students put various tasks for the Holocaust and the contemporary witness reports of survivors. Among other things, the chatbot should summarize five personal reports, which were recorded in La Paz, Krakow and Connecticut by Holocaust survivors in 1995.
The result: The AI reported that those affected had suffered greatly. However, she failed to make emotional details of the stories and thus caused the extent of suffering. For example, Chatgpt did not mention that the mother of the then seven -year -old Luisa D. in 1942 on the run from the Nazis, to keep her daughter from thirsting through the moisture of her blood. “This intimate, most frightening moment – unspeakable and unforgettable – was made invisible by a model that is trained to put the likelihood about what is profound,” writes Burzlaff.
The personal horror of Luisa D. was reduced to a negligible event among many – because it does not fit into a known category. However, since human experiences are never uniform and do not necessarily follow a pattern, algorithms cannot weight or classify them properly, says Burzlaff. The result is a decoupled and shallow copy of reality because the essential and outstanding lack. The AI does not manage to understand the complex meaning of the events under the words and to classify their emotional and ethical value.
What are the tasks of historians?
According to Burzlaff, such failures of artificial intelligence prove that human historians remain indispensable in the AI age in order not to let the historiography to have a superficial, smoothed overview. Only people are able to capture and capture the emotional and moral aspects of historical events in all their complexity. This applies not only to the extreme example of the Holocaust, but also to less serious world events. “The challenge that lies ahead of us is not to reject AI, but to reaffirm what historical thinking actually requires: contradictions, gaps, positionality and the slow work of finding meaning,” says Burzlaff.
According to the researcher, historians should therefore not rely too much on the support of KIS so as not to lose sight of the true importance of their work. Instead of evaluating large amounts of archives from archives with algorithms and AI summary, you should rather concentrate on the interpretation and importance of events. A differentiated approach is particularly important when it comes to trauma, genocides and historical injustices.
“Historians have to be aware of what these systems can do and what are not,” writes Burzlaff. “They summarize, but do not listen. They reproduce, but do not interpret. They are characterized by coherence, but fail due to contradictions.” In his article, Burzlaff wrote corresponding guidelines how teachers and researchers should deal with AI and historical topics. This includes doing exactly what AI cannot do: “pause, return and interpret”.
Source: Cornell University; Specialist articles: Rethinking History, DOI: 10.1080/13642529.2025.2546174