AI Challenges Holocaust Memory by Losing Context, Study Finds
Translated & summarized from Ynet by baba
A Tel Aviv University study by Dr. Yael Cohen argues that AI tools struggle to preserve the full context of Holocaust testimonies, potentially losing crucial meaning. Examining three archival items, the research highlights how physical characteristics like stains and carvings are integral to the testimony. The study concludes that responsible digital archiving must maintain the link between transcriptions and original artifacts. AI can provide access to testimonies, but preserving the material context is essential for complete understanding.
The story in 6 lines · by baba
- AI tools risk losing crucial context in Holocaust testimonies, according to a Tel Aviv University study.
- Physical characteristics of artifacts, like stains and carvings, are integral to Holocaust testimonies.
- The study examined clay tablets, an illustrated poem, and a cloth document from the Holocaust era.
- AI struggled to distinguish between text and material degradation in the analyzed items.
- Responsible digital archiving must preserve the link between transcriptions and original artifacts.
- Dr. Yael Cohen stated that more readable text is not necessarily more complete testimony.
A new study by Dr. Yael Cohen from Tel Aviv University's English and American Studies department explores the challenges artificial intelligence faces in preserving the full context of Holocaust testimonies. The research, published in the journal AI & Society, examines three complex archival items from the Holocaust era, arguing that physical characteristics like stains, folds, carvings, and burn marks are integral parts of the testimony itself.
The study involved digitally processing images of three archival items, including text recognition, correction, translation, and cataloging. The analysis assessed AI tools' ability to extract information while also identifying the meanings that could be lost in the process. The selected items were chosen for their complexity, not as a representative sample of all Holocaust writings.
One case study involved two clay tablets on which writer Eliezer Heiman carved his family's names and messages for future generations while hiding near Kovno in September 1943. The study noted that AI tools struggled to distinguish between the carved letters and material degradation, shadows, or soot marks. However, these marks are significant, relating to the destruction of the hiding place and the tablets' survival.
Another example is an illustrated poem by Eva Schulze, created in the Theresienstadt Ghetto in 1943, depicting ghetto life and longing for home. The final lines, written in faint pencil, were particularly difficult for AI to identify. The study found that even when words were successfully extracted, the connection between the poem and its accompanying illustrations weakened in the digital output.
The third item was a piece of cloth on which Sal Barmson, a young Jewish man from Amsterdam, documented his prison cell with drawings and explanations in January 1943. This cloth was smuggled to his mother. The research suggests that separating the text from the drawings and the fabric itself could obscure the item's nature as a personal message conveying his conditions in captivity.
Dr. Cohen concludes that responsible digital archiving must maintain the link between transcriptions and original artifacts, document alterations, and flag uncertain readings. She emphasizes that AI can provide new generations access to Holocaust testimonies, but the responsibility extends beyond deciphering words. The physical traces, the context, and the material nature of these artifacts are crucial for conveying their full meaning. "More readable text is not necessarily a more complete testimony," she stated. Cohen added that the study highlights the importance of fragility, ambiguity, and context of these rare objects, especially in the age of AI, hoping it will lead to AI tools that better account for contextual factors to ensure the human voice continues to be heard.