Imagine finding a single page from a medieval book and not knowing where the rest went. Another piece might be held by a library in a different country, perhaps catalogued under an unrelated description. Cambridge researchers are developing AI manuscript matching to help scholars identify fragments that could once have belonged to the same volume. The ambition is not to let a machine rewrite history, but to make a daunting search more manageable.
The project is part of a new set of AI-deas Sprints announced in October 2026. It will focus initially on the Cairo Genizah, a vast collection of medieval Hebrew and Judeo-Arabic manuscript fragments. Researchers will use a system to rank likely matches between images, then examine the suggestions themselves. The process is deliberately collaborative: human decisions become part of refining the tool.
Why matching a medieval page is like a huge jigsaw puzzle
A torn page can carry clues about the book from which it came. The script may resemble that of another fragment; ruling lines, damage, material and the shape of an edge may help an expert decide whether two pieces belong together. Yet a scholar often has to search across collections with different catalogues and digitisation histories.
The difficulty is not simply that the pieces are numerous. The fragments can be incomplete, photographed under different conditions or separated from the contextual evidence that once identified them. A pair of similar-looking texts could have been copied by different scribes; two genuinely related pieces may have suffered different forms of wear and damage.
Cambridge says the Cairo Genizah contains around 350,000 medieval Hebrew and Judeo-Arabic fragments, with approximately 60% held at Cambridge University Library. These figures show why manual comparisons are so demanding. Even when a researcher knows the language and historical period, the number of possible combinations is enormous.
Digital images allow scholars in different places to inspect material without constantly moving fragile originals. But the arrival of images creates another problem: finding the relevant items in a large digital collection. A search tool that prioritises plausible matches could reduce the amount of repetitive comparison needed before an expert can make a judgement.
How AI manuscript matching is meant to work
The project description says a model will analyse pictures of manuscript fragments and rank those that appear likely to come from the same original manuscript. Scholars will inspect the proposals, confirm or reject them and document their conclusions in a shared workspace. Their decisions are intended to improve later recommendations.
The ranking is important. A system need not declare that two fragments definitively match to be useful. It can offer a shorter list of candidates that deserve attention. Researchers can then bring historical knowledge, language skills and close examination to bear on the possibilities, including evidence that an image alone may not capture.
The team is led by Vincent Beiler of Cambridge University Library and John Dudley of the Department of Engineering. The sprint is intended to scale an existing prototype and develop a collaborative platform. That means the announcement is about building and testing a more capable research workflow, not a claim that the Genizah has already been digitally reconstructed.
There is a subtle distinction between finding resemblance and establishing a relationship. Pages from the same period or writing tradition may share visual characteristics without ever belonging to the same book. A confident model score is therefore a suggestion worth investigating, not a substitute for archival judgement.
Even a rejected match can be useful if the system makes its reasoning inspectable. A scholar might recognise that two pieces use similar paper or lettering but belong to different textual traditions. Recording that distinction could help avoid repeatedly investigating the same false lead. It could also expose where the model is overconfident, allowing engineers and historians to refine future searches together rather than operating as separate teams with little feedback.
What lost books can tell us when their pages return
Reuniting fragments could help scholars recover the order of a text, identify a previously unknown work or clarify how a document was copied and used. A page that looks ordinary in isolation may become more informative when placed beside related material. The value lies in the evidence about people and places, not simply in completing a digital puzzle.
Medieval manuscripts are also records of ordinary life. Beyond famous works of literature and religion, surviving material can illuminate trade, correspondence, education and the movement of ideas. Bringing related fragments together can change how a scholar interprets individual passages, even when the pages themselves remain physically scattered.
That makes the project different from AI-assisted work on sealed ancient scrolls. In one case researchers try to recover writing from material that is difficult to open or read. In this case the writing may already be visible, but the relationships between dispersed fragments are missing. Both depend on careful interpretation rather than treating a model’s output as historical fact.
The reconstructed manuscript will usually be a scholarly claim supported by images and evidence. It may remain open to revision. A useful collaborative tool should preserve the reasons for each proposed match and make disagreement possible, because historical scholarship advances partly through the ability to question earlier conclusions.
Why a human must still make the final connection
Historical material brings unusual challenges for AI. Digitised pages vary in colour, contrast, scale and condition. Scripts change across places and generations. A stain may obscure the most important mark, while an unusual letter form might distinguish a scribe. A system trained on readily available images could struggle with rare or damaged examples.
Bias can also enter through the collections that have been photographed and catalogued. If one library has better images or richer records, the model may learn more from that material than from fragments held elsewhere. A match-ranking service should make those limitations visible instead of creating a false impression of complete coverage.
The human element is not a decorative safeguard. Specialists can read language, assess writing materials, inspect context and weigh explanations that may not be evident to image software. A good digital workspace would help them compare findings and record uncertainty, rather than hide expert disagreement behind a single automated answer.
Similar questions surround AI use in archaeological interpretation. Computer systems can highlight patterns that deserve inspection, but a pattern is not the same as an explanation. Researchers still need independent evidence and a defensible account of what the discovered material means.
There is an intriguing practical benefit in making the reconstruction work collaborative. A scholar on one continent may recognise a letterform while another understands the history of a collection where a possible matching fragment is held. A shared record could allow both contributions to be preserved without implying either researcher has surrendered judgement to the software. Good documentation would make a proposed reconstruction easier for future specialists to inspect or dispute.
The images also need careful stewardship. Scholars may work from digitised surrogates, but the institutions holding the originals retain responsibilities for conservation, catalogue quality and rights. A technically possible comparison is not automatically permission to republish every image without conditions. The programme will have to fit the workflows and policies of participating collections as well as demonstrate a useful algorithm.
What the Cambridge team will have to prove
The sprint programme gives teams a period to develop and test their ideas with the people who could use them. For manuscript matching, success would not be measured merely by how many possible pairs the software proposes. It would involve the quality of those suggestions, the time scholars save, the diversity of material the system can handle and the usefulness of the collaborative records.
Independent evaluation should examine the false matches as carefully as the successful ones. A tool that returns a very long list of weak candidates could create more work, not less. A system that ignores unusual documents might quietly steer attention away from the fragments most in need of specialist investigation.
This is part of the wider promise of digitising cultural heritage. Making an archive searchable is valuable, but discovering meaningful connections across it requires a further layer of organisation and analysis. AI can assist with that layer if the results remain transparent and open to correction.
No dramatic reunited volume should be announced before the evidence supports one. The more realistic story is that a new tool may help scholars spend less time searching blindly and more time reasoning about manuscripts that matter. For fragments separated across the world, that could be an important advance even if the most exciting discoveries are still to come.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.