发生了什么
BBC Future reports that researchers are using machine learning to accelerate the transcription and decoding of handwritten historical ciphers. A team associated with the University of Stockholm helped decode the 408-page Borg cipher, while researchers in the Descrypt project are developing systems that can process unusual symbols, historical handwriting and unknown scripts.
BBC Future reports that the Borg cipher, a 408-page manuscript held by the Vatican Library, was written with 34 unusual symbols, some Roman letters and an Arabic title page. According to the report, researchers used machine learning to help identify the underlying substitution cipher and found thousands of purported treatments, including remedies involving red wine and fermented nutmeg.
The manuscript is available online through the Vatican Library under the reference Borg.lat.898, according to the BBC. The report does not independently establish the historical accuracy or medical value of those treatments. The BBC describes the work as a combination of several tasks that have traditionally required substantial human effort.
Researchers first transcribe handwritten symbols into machine-readable form, then use cryptanalysis to infer how the symbols correspond to letters or words. Existing handwriting systems can handle some historical languages and writing styles, but the report says they often struggle with invented symbols, astrological signs, unusual numbers, faded ink and poor handwriting. A letter sent in 1637 by Sigismund Heusner von Wandersleben was transcribed with help from the AI platform Transkribus, although researchers still had to make manual corrections.
Researchers involved in the Descrypt project are developing a system intended to combine transcription and deciphering in one process. BBC Future reports that they tested an approach on the 105-page Copiale cipher, a decoded German manuscript describing the rituals and rules of an 18th-century secret society. The system was trained using generic handwriting, images of cipher lines and corresponding decoded German text, then accurately decoded portions it had not previously seen, according to the report. The team also built a chatbot-like tool combining cipher algorithms, historical language models and image-recognition systems. In a reported test on a 500-symbol extract from the Borg cipher, the tool translated and decoded the passage in just over 29 minutes and documented why its proposed solution appeared plausible.
为什么这很重要
The work could make previously inaccessible archival material easier for historians and the public to study. The BBC says the systems may help investigate diplomatic messages, medical knowledge, religious practices and personal correspondence that have remained unread because of difficult handwriting, damaged pages or unfamiliar encryption.
The practical significance is that archives contain large quantities of material that historians cannot easily search or interpret. BBC Future cites estimates that roughly 1% of material in libraries and archives worldwide is partly or wholly encrypted. Even if that estimate is uncertain, the underlying problem is concrete: documents may use substitution systems, multiple symbols for one letter, deliberate decoys or languages that are no longer understood. Manual work can be slow; the BBC reports that deciphering a three-page letter from Holy Roman Emperor Charles V took six months because it used 120 different cipher symbols.
AI-assisted decoding could expand the range of historical evidence available for research. The BBC says encrypted documents may contain diplomatic correspondence, medical practices, rituals, romances and everyday details omitted from established historical narratives. It points to letters written by Mary, Queen of Scots, whose decoded contents reportedly shed light on alleged plots to regain her throne and on her relationship with her son. The report also describes a letter from the Thirty Years’ War whose decoded text contained warnings about suspected conspiracies among Swedish Protestant allies.
These examples show why decoding is not merely a technical exercise: the result can affect interpretations of people and periods. The reported approach may also matter where conventional language assumptions fail. Large language models are usually trained on enormous collections of readable text, while historical cipher datasets are much smaller and harder to assemble. Descrypt researchers are therefore compiling encrypted archival material, including 400 coded postcards from the late 19th and early 20th centuries. BBC Future reports that some decoded fragments appear to be German love letters.
A system that can learn from paired cipher symbols and their decoded forms, while using historical language knowledge as a source of clues, could help prioritize archival work. But the source does not establish that the tool is broadly available, peer-reviewed, or ready for unsupervised historical interpretation.
互动机制：它实际上是如何运作的
以交互方式探索这一发展背后的基础技术。
A route planner searches possible journeys using explicit rules. What does this illustrate about AI?
接下来看什么
The important tests will be whether these systems work reliably on documents whose answers are unknown, whether they can distinguish plausible readings from fabricated ones, and whether researchers can assemble enough representative historical data for training. The BBC’s reported results have not been independently confirmed from primary research materials here.
The central issue is verification. A cipher can produce a grammatically convincing but incorrect reading, especially when the underlying language, symbol system or historical context is uncertain. The BBC reports that the Descrypt tool records its reasoning and explains why a solution is plausible, which researchers see as a way to reduce hallucinated interpretations. That documentation is useful, but the source does not provide independent accuracy rates, confidence thresholds, results across unknown ciphers, or examples where the system rejected an attractive but wrong solution.
Future evaluations should test documents whose contents remain unknown rather than only manuscripts that researchers have already deciphered. The reported Copiale and Borg experiments benefit from existing reference material, which makes them useful demonstrations but does not by itself prove that the system can solve genuinely unsolved texts. It is also unclear how performance changes with damaged pages, rare writing systems, multiple encryption layers, deliberate decoy symbols or a language absent from the training data.
The BBC says the team hopes to address some of these challenges by expanding training across scripts, alphabets and symbolic repertoires. Availability and governance are additional unknowns. BBC Future describes a research tool that researchers or members of the public might eventually use, but it does not say that the combined chatbot is publicly released or specify its access terms. Wider use could help libraries and independent researchers, while mistaken machine readings could enter databases or popular histories before experts review them.
This assessment relies only on the BBC report; the cited technical findings, timing claims and historical interpretations have not been independently confirmed here. The next meaningful development would be a public system or research paper reporting reproducible tests, error analysis and expert validation on previously undeciphered material.