History is inherently "lossy." Over millennia, the physical record of human civilization—whether inscribed on clay, parchment, or stone—has been eroded by time, conflict, and environmental decay. These missing segments, known as lacunae, create fundamental voids in our understanding of the past. Traditionally, scholars like Thomas Gärtner have relied on rigorous philological analysis and "expert intuition" to propose restorations for these gaps.
However, a new research direction is reframing historical reconstruction as a computational problem. By treating damaged archives as "corrupted data" and history as a sequence restoration task, researchers are using generative AI and deep learning to bridge the gaps between surviving fragments.
History as a Masked Language Task
The most intuitive breakthrough in this field treats ancient text restoration like a modern NLP task. In models like BERT, "masked language modeling" involves hiding words in a sentence and asking the model to predict them based on the surrounding context. Researchers have realized that a 3,000-year-old broken cuneiform tablet is essentially a "naturally masked" sequence.
Koren Lazar and colleagues demonstrated this by applying BERT-style architectures to Ancient Akkadian texts. Because these models are trained on vast corpora of similar inscriptions, they can "hallucinate" the missing signs with high statistical probability. Similarly, the Logion framework, developed by Charlie Cowen-Breen and others, uses this logic to assist philologists in editing premodern Greek manuscripts, identifying scribal errors and filling gaps that have persisted for centuries.
Visual and Structural Hallucination
Restoration isn't limited to text; it extends to the physical and visual integrity of artifacts. When a document suffers from "ink rot" or physical holes, the semantic meaning and the visual texture are lost simultaneously.
The Historical Document Repair (HDR) task uses diffusion-based models to perform "image-to-image" translation. Unlike simple noise reduction, these models use semantic priors to reconstruct entirely missing characters and textures. This approach is mirrored in the study of architectural heritage, where Yifan Li and team use multimodal learning to reconstruct lost Traditional Chinese Gardens from sparse textual descriptions, turning 2D historical narratives into 3D spatial models.
Even when the text is entirely illegible, the "silhouette" of the medium itself carries data. Danielle Kapon and Michael Fire have shown that Variational Auto-Encoders (VAEs) can date cuneiform tablets based solely on their 2D physical shape, providing a "proxy provenance" when the written record fails.
Reassembling the Global Puzzle
One of the greatest barriers to reconstruction is the physical dispersion of fragments across global museums. A single Babylonian tablet might exist as three pieces in London, Berlin, and Baghdad.
Computational methods are now being used to find these "joins." By aggregating thousands of transliterated tablets into platforms like the Electronic Babylonian Library, researchers like Enrique Jiménez have successfully identified over 1,200 "joins," effectively reuniting physically separated artifacts in digital space. This process transforms a collection of disconnected data points into a cohesive narrative, such as the recovery of previously unknown Babylonian hymns.
Critical Confabulation and Social Gaps
Perhaps the most provocative application of this technology is "critical confabulation." While engineers usually try to eliminate LLM hallucinations, Peiqi Sui and Eamon Duede argue that hallucination can be a feature for social good.
Historical archives are not just physically damaged; they are systemically biased, often omitting the stories of marginalized individuals. By using models to speculate on these "hidden figures" based on surrounding social and political contexts, researchers can fill "archival silences" that were created by institutional neglect rather than physical erosion.
The Shift Toward Synthetic History
The common thread across these 87 papers is a shift in the role of the historian and the archivist. We are moving from a paradigm of preservation (keeping what remains) to one of synthesis (reconstructing what was lost).
Whether it is recovering 300 years of climate data from royal diaries or predicting the rules of ancient games from broken boards, the goal is the same: using computational probability to overcome the "irrecoverable gaps" noted by Richard E. Leakey decades ago. By treating the void not as an end-point, but as a prompt, this field is slowly turning the "fragmentary record" into a complete, albeit synthetic, history.