History education research on AI is somewhat like the race between the hare and the hedgehog: in the one or two years between designing a study and publishing its findings, many empirical results are already overtaken by rapid technological developments. Models become more powerful, agents and multimodal systems emerge, applications become more accessible, and previously poorly understood functionalities quickly turn into common knowledge. And yet, history education must engage with the empirical study of AI with particular urgency. The transformations associated with these technologies are highly relevant for historical culture and learning: from the pollution of the digital space through ‘AI slop’ and the deepfaking of primary sources to the perpetuation of biases and the streamlining of the collective historical consciousness. It is therefore essential to better understand how AI works and to integrate these insights into appropriate teaching and learning arrangements.

In light of this, my study explored how AI image generators produce representations of the past and what didactic value can be derived from them. At its core lies the question of what these visualisations, produced within seconds, can achieve from a historical perspective: how authentic and plausible they appear; which symbols, stereotypes and narrative patterns they rely on; which implicit datasets and biases become visible; and whether they are capable of applying higher-order thinking skills or merely reproducing their training data.

An analysis of a dataset of 1,024 images from five leading models reveals an ambivalent picture. While the systems are capable of producing convincing epochal atmospheres, they lack concrete historical referents and independent judgement when it comes to assigning significance or making historical comparisons. Instead, stereotypical and often heavily biased visual motifs dominate, frequently marked by misattributions. Yet it is precisely these “errors” that are didactically productive: they create opportunities to confront students with anachronisms, perspectivity, bias, and questionable historical judgement, thereby fostering historical thinking. At the same time, they offer valuable insights for research on historical culture, as they make dominant patterns of interpretation and judgement visible.

In the fable mentioned at the outset, the hedgehog ultimately wins the race not through speed, but through cunning. Similarly, history education research on AI does not need to keep pace with every technological development in detail. What matters instead is making broader patterns visible: typical functional logics, recurring structures and central problem areas, as well as how students tend to engage with this emerging technology. Publishing in HERJ provides an opportunity to situate these findings within a wider subject-specific discussion and to make them accessible for both research and practice.

Digital déjà vus: an investigation into how AI image generators s(t)imulate historical thinking by Heinrich Ammerer (University of Salzburg, Austria) is part of the HERJ series Investigating the Relationship Between Generative Artificial Intelligence and History Education and published in History Education Research Journal, volume 23

Heinrich Ammerer, PD MMag. Dr., is an Associate Professor of History and Civic Education at the University of Salzburg (Austria) and a history teacher. His research focuses on the psychogenesis of historical consciousness, conceptual learning and the use of audiovisual media in the classroom.

AI artificial intelligence image generation Midjourney DALL-E stable diffusion historical significance authenticity historical unconscious