High-Risk Memories? Comparative audit of the representation of Second World War atrocities in Ukraine by generative AI applications

A team of researchers led by Mykola Makhortykh conducted a comparative audit of three major generative AI applications, examining how they handle representations of Second World War atrocities in Ukraine. Published on arXiv as "High-Risk Memories?", the study focuses on what they term "high-risk memories"—emotionally charged historical events that are frequently instrumentalized in political discourse. The researchers systematically tested the models for various forms of misrepresentation, ranging from outright factual errors (hallucinations) to more subtle issues like inconsistent moral framing of events and selective depiction of victim groups.

The audit's findings reveal that generative AI poses a substantive risk of historical distortion. The models demonstrated patterns of "selective moralization," where the moral weight assigned to similar atrocities varied inconsistently based on contextual prompts. This inconsistency, combined with factual inaccuracies, suggests that AI-generated content could significantly alter public understanding and memory of these contested events. The researchers argue that as AI becomes a primary tool for information discovery and content creation, these distortions could accelerate and cement biased historical narratives, with particular danger for regions like Ukraine where historical memory is actively contested.