What AI can lose when it makes Holocaust writings easier to read

In an Amsterdam prison during the Nazi occupation, Sal Bramson drew his cell on pieces of cotton cloth. He added handwritten explanations and sent these messages to his mother, concealed inside a shirt collar among the prison laundry. The cloth could travel because it passed as something ordinary. Its material was part of what made communication possible.

I study Holocaust writings that survive on fragile or improvised surfaces. In these objects, the words and the things carrying them remain intimately connected. Bramson's captions guide his mother's eye around his cell. His drawings give shape to the space he describes. The small, crowded textile carries both his account of confinement and evidence of his effort to reach her.

What happens to these relationships when an archival object becomes data?

AI and other computational tools can make historical collections easier to search, transcribe and translate. These are substantial gains. A person unable to decipher handwriting or read the original language may encounter a document that would otherwise remain inaccessible. Yet the process of making words available can also obscure the material evidence through which they survive.

My recent study, published in AI & Society, examines this problem through a concept I call "material uncertainty." It describes writing whose meaning remains bound to the damaged, improvised or unstable form in which it reaches us. The uncertainty may concern the circumstances of composition, the surface carrying the words, their intended recipient or the route by which the object entered an archive. These relationships shape what the writing allows us to know, including what remains unresolved.

I followed digital reproductions of three Holocaust writings through a sequence of computational operations: enhancing images, separating regions, recognizing handwriting, correcting and translating text, and generating descriptions for searchable records. I examined what each operation preserved or weakened. I used these deliberately difficult cases to trace how particular workflows change an object's representation. The study makes no claim to measure AI's overall accuracy.

Bramson's cloth showed how an apparently helpful operation could complicate interpretation. Suppressing the visible weave made some marks easier to distinguish. Separating drawings from writing helped organize the image for processing. But the textile was the means of concealment, and the captions and drawings worked together to address his mother. Handwriting recognition also produced fragments that correction could make plausible without establishing what Bramson had actually written.

A second object was an illustrated pencil poem by Eva Schulzová, a child imprisoned in Terezín. Her page moves between the camp's exterior, the women's bunks and an imagined meal of turkey and cake. The placement of words and drawings participates in that movement. She was later deported to Auschwitz-Birkenau, where she was murdered.

The workflow recovered much of her verbal text. Yet its generated description listed the illustrations without adequately representing their relationship to the poem. It described the imagined meal as a "food still life." I read that imagined abundance in relation to the child's confinement and hunger. The label gives little sense of that relationship. A searchable description can be informative and still leave the reader far from the work it describes.

Eliezer Heiman's paired clay tablets posed a different difficulty. Inscribed in Hebrew while he was in hiding in the Kovno ghetto, they record names and address future generations. Their paired form recalls the biblical Tablets of the Law. The surviving surface bears smoke staining and flaking, while the letters are cut into the clay.

In a photograph, shadows cast by engraved letters and damaged areas can look alike. The workflow could enhance some features but could not reliably separate inscription from deterioration. Here, the problem began with how the object was photographed. Imaging it under multiple lighting directions could better record its relief.

That distinction matters for anyone building an archive. Some evidence is missing because the initial image never captured it. Other evidence is visible in the image but becomes less accessible when processing separates elements that belong together. These problems require different responses.

I argue for archival systems that keep readers connected to the surviving object throughout these transformations. A transcription should link back to the relevant area of the original image. Readers should be able to see what was enhanced, suppressed, inferred or manually corrected. Drawings and captions should remain connected. An uncertain reading should remain identifiable as uncertain, even when a fluent translation is available.

This approach also gives scholars room to clarify what can responsibly be clarified. A stain may have several possible causes; its significance needs investigation. Preserving uncertainty means making the evidence and the limits of an interpretation available for scrutiny.

The question extends beyond Holocaust collections to how libraries and museums present fragile historical records. When a polished transcription circulates on its own, readers may have little reason to ask what happened between the object and the screen.

I return to Bramson's cloth because it asks us to keep two journeys in view: its passage out of prison to his mother, and its passage through our technologies to us. I want computational access to bring readers closer to his words while keeping the first journey visible.

This story is part of Science X Dialog, where researchers can report findings from their published research articles. Visit this page for information about Science X Dialog and how to participate.

Yael S. Hacohen is a postdoctoral fellow in the Department of English Literature and American Studies at Tel Aviv University. Her research examines Holocaust literature and testimony, poetry, and the material forms through which writing survives. Her current project, Last Poems of the Holocaust, brings together archival research and literary interpretation.