Colonial Bias in AI Training Data: Prompting Sora to Generate Images of Aotearoa New Zealand's Historical Past

This paper examines how generative artificial intelligence (AI) reproduces colonial visual tropes when tasked with representing Aotearoa New Zealand's historical past. Using OpenAI's Sora as a case study, the analysis investigates AI‐generated images prompted to depict (1) precolonial landscapes, (2) first contact between Māori and Europeans, (3) British colonial rule, and (4) Māori figures from the 1860s. Drawing on iconographic methods, the study finds that Sora‐generated outputs closely mirror dominant settler‐colonial visual conventions. These include portrayals of the land as terra nullius, colonisation as peaceful and consensual, and Māori as timeless, passive figures. Rather than offering disruptive alternatives, Sora reinforces hegemonic memory frameworks learned from biased training data. As generative AI tools become increasingly influential in shaping public understandings of the past, such depictions matter; they naturalise myths of benevolent colonisation and undermine Māori claims to political sovereignty, redress, and cultural revitalisation. The paper concludes by evaluating possible interventions at three stages of the AI development pipeline—preprocessing, model training, and postprocessing—while also highlighting the importance of AI literacy in enabling users to critically prompt and repurpose these technologies for decolonial ends.