Researchers say AI fails to describe complexities of Holocaust

Richardson-Walden pointed out that GenAI systems operate as “knowledge machines,” assigning probabilistic values to words and phrases without considering their cultural and historical significance, resulting in the overshadowing of lesser-known facts and narratives in favor of well-known, often sensationalized stories.

She highlighted the limitations of systems like ChatGPT in summarizing the intricate histories of events like the Holocaust, noting that attempts to condense complex events spanning multiple years and locations into brief summaries are inherently problematic.

While the research does not offer definitive solutions to these challenges, Richardson-Walden is exploring alternative approaches in collaboration with informatics and engineering experts to address the difficulty of encoding cultural nuances into AI training data.

The lab’s policy briefing also raised concerns about the inadequate data present in commercial GenAI models, particularly when dealing with sensitive historical subjects such as genocide and conflict, underlining the importance of accurate and standardized metadata in digitized Holocaust materials.

Another issue highlighted in the briefing is the self-censorship embedded in many commercial image GenAI models, which often refuse to generate images related to the Holocaust, leading to limitations in visual representations of historical events.

Richardson-Walden emphasized the need for a balanced approach in implementing censorship guidelines within AI programming that prevents misinformation without erasing crucial historical narratives, advocating for open dialogue among tech companies, academic institutions, and Holocaust memory experts to navigate these complexities.

The Landecker lab’s website references the USC Shoah Foundation’s Dimensions in Testimony AI model as a notable example of Holocaust memory digitization, showcasing a domain-specific GenAI model that incorporates human input to provide interactive testimonies from survivors and experts.

However, not all organizations have the resources to develop such specialized models, underscoring the importance of widespread digitization efforts to support responsible utilization of AI in preserving Holocaust memory.