Existing AI models don’t capture the full depth and complexity of the Holocaust, offering instead simplified narratives, warns the Landecker Digital Memory Lab at the University of Sussex. Launched in November 2024, the lab aims to secure a sustainable future for Holocaust memory and education in the digital era.
In a recent briefing to the International Holocaust Remembrance Alliance, the lab pointed out the issues with using mainstream AI, like generative AI systems such as ChatGPT and Gemini. These models struggle because they lack accurate, comprehensive data about the Holocaust and don’t include the insights of experts in this field. Victoria Grace Richardson-Walden, the lab’s lead investigator, called on stakeholders in Holocaust memory and education to digitize their data and expertise. Instead of just encouraging visits to museums and memorials, they should develop clear digitization strategies. “Very few have a solid plan,” she noted, referencing archives, museums, and libraries worldwide that only digitize materials for specific exhibitions.
Richardson-Walden also raised concerns about the risks to cultural heritage amid ongoing conflicts, like those in Ukraine and the Middle East. She emphasized that when history gets distorted for political purposes, especially on social media, it loses its nuance.
She criticized generative AI models for not being true “knowledge machines.” They treat words as mere data points, which can obscure less-known facts by focusing on popular narratives. “You can’t condense six years of history affecting countless people into a few bullet points,” she said.
While the lab aims to explore alternatives with tech experts, they acknowledge that training AI on nuanced historical content is challenging. “Cultural elements are tough to code accurately,” Richardson-Walden explained.
Furthermore, she highlighted the issue of unreliable data in commercial generative AI, particularly regarding sensitive subjects like genocide. “For AI to work well, Holocaust organizations need to digitize their materials strategically, ensuring correct metadata,” she stated.
Self-censorship in many AI models also hinders learning. When asked for Holocaust images, users often face censorship. For instance, DALL-E can only produce very generic images like wreaths or vague representations. Richardson-Walden pointed out that while this censorship aims to prevent misinformation, it inadvertently makes the Holocaust less visible. Developers must strike a balance: protect against false narratives while not restricting education.
“Dialogue is key,” she said, advocating for discussions between tech companies and organizations involved in Holocaust memory. The Landecker Lab even offers free consulting for tech firms new to this challenging area.
The lab cites the AI model Dimensions in Testimony from the USC Shoah Foundation as an example of effective Holocaust memory digitization. This model blends AI with extensive human oversight, allowing interaction based on survivor testimonies. However, not all memory centers have the resources Landecker does, so focusing on mass digitization is crucial for responsibly informing larger AI systems.