Charred manuscripts from a villa near Mount Vesuvius have resisted researchers for centuries. Today, artificial intelligence is reconstructing their content, revealing unknown works by ancient philosophers. How is this possible, and why could this breakthrough change our knowledge of antiquity?
Charred legacy: Why were the Herculaneum scrolls unreadable?
In 79 AD, the eruption of Mount Vesuvius buried not only Pompeii but also nearby Herculaneum under a layer of ash. Among the ruins of a villa now known as the Villa dei Papiri, over 1,800 scrolls were discovered in the 18th century – the only surviving ancient library. Unfortunately, high temperatures and a lack of oxygen caused the papyri to carbonize, turning them into brittle lumps of charcoal. Attempts to unroll them resulted in destruction – many scrolls fell apart in the hands of researchers.
For over 250 years, scientists unsuccessfully tried to read their content. Chemical and mechanical methods were used, as well as attempts to "evaporate" layers using lasers. All these techniques proved too invasive. Only the development of X-ray imaging and machine learning opened a new era in the study of these unique documents.
Vesuvius Challenge: How AI became the key to antiquity
In 2023, the Vesuvius Challenge initiative offered a million-dollar prize for reading text from the scrolls. The project joined the forces of scientists from the University of Kentucky, Oxford, and ETH Zurich with artificial intelligence experts. The application of two technologies proved crucial:
- Phase-contrast X-ray tomography – allowed for the creation of 3D scans of the scrolls' interior without physically unrolling them. Carbon-based ink, though barely visible, differs in density from carbonized papyrus.
- Machine learning algorithms – neural networks trained on simulated data learned to recognize letter shapes in 3D scans. The greatest success was achieved by U-Net models, which specialize in medical image segmentation.
In October 2023, student Luke Farritor read the first word: "πορφύρασ" ("purple"). A few months later, the team of Youssef Nader and Julian Schilliger revealed the first full sentences, likely authored by Philodemus of Gadara – an Epicurean philosopher from the 1st century BC.
What do the scrolls hide? First discoveries and their significance
The deciphered fragments shed new light on Epicurean philosophy. One of the texts discusses the role of music in human life:
"Regarding music, some say it provides no benefit. But does it not make life more pleasant?"
Other fragments mention feasts, wine, and wealth – topics typical of Philodemus, whose works were previously known to us mainly from fragmentary quotes. If the authorship is confirmed, the Herculaneum scrolls could become the most important source for the study of Epicureanism in Rome.
Interestingly, these texts are not copies of known works, but previously unknown writings. This means we are dealing with new source material that could change our understanding of ancient philosophy. As Prof. Richard Janko of the University of Michigan notes: "It's like finding a lost book of the Bible – only for the history of philosophy."
How does the algorithm reading burnt manuscripts work?
The process of reading the scrolls is a combination of precise physics and advanced computer science. Here is how it looks step by step:
- 3D scanning: The scroll is placed in an X-ray tomograph, which takes thousands of images from different angles. Phase-contrast technology allows for capturing subtle density differences between the ink and the papyrus.
- Virtual reconstruction: Algorithms combine the images into a 3D model that can be "sliced" into any layer – like a virtual bread slicer.
- Segmentation: Neural networks analyze the model, identifying places where ink is deposited on the papyrus surface. This is the most difficult stage – the ink is often invisible to the naked eye even in scans.
- Letter recognition: Models trained on synthetic data (simulated scans of scrolls with letters applied) attempt to match shapes to known Greek alphabets.
- Text reconstruction: Algorithms combine letter fragments into words and sentences using linguistic context. Here, philologists assist by verifying the accuracy of the reading.
The biggest challenge is the lack of training data. Since there are no "perfect" scrolls for comparison, scientists must rely on simulations. As project leader Dr. Brent Seales says: "It's like teaching a child to read by only showing them pictures of blurry letters."
Limitations and pitfalls of artificial intelligence in archaeology
Although the success of the Vesuvius Challenge is impressive, the method is not without flaws. The biggest risks are:
- False positives: The algorithm might "see" letters where there are none – for example, in papyrus cracks or accidental contamination.
- Interpretative errors: Ancient Greek has many dialects and spelling forms. AI might incorrectly combine letter fragments, creating non-existent words.
- Linguistic distortions: The algorithm does not understand cultural context. For example, it might interpret an abbreviation as a full word, changing the meaning of the text.
To minimize errors, scientists use multi-stage verification:
- Comparison of deciphered fragments with known texts by Philodemus and other authors of the era.
- Independent analysis by different teams – if two algorithms agree on a reading, the probability of error decreases.
- Consultations with classical philologists who assess the linguistic and logical correctness of the reconstruction.
Despite these precautions, some researchers emphasize that the final interpretation of the text will always be a task for humans. As Dr. Roberta Mazza of the University of Manchester notes: "AI can read the letters, but it won't understand what they mean in the context of ancient culture."
Is this the beginning of a revolution in archaeology?
The success of the Herculaneum scrolls opens new possibilities for the study of damaged texts. Similar projects are already underway in other parts of the world:
- Dead Sea Scrolls: The Leon Levy Dead Sea Scrolls Digital Library project uses AI to reconstruct fragments and identify scribes.
- Sinai Palimpsests: As part of the Sinai Palimpsests Project, algorithms read texts hidden under later writings on parchment.
- Pompeii inscriptions: A team from the University of Naples is using machine learning to read faded inscriptions on the city's walls.
Interestingly, this technology could also find applications outside of archaeology. For example:
- Reading damaged archival documents (e.g., court records, old maps).
- Analyzing damaged digital media (e.g., old floppy disks or magnetic tapes).
- Reconstructing texts in forensics (e.g., burnt letters or documents).
However, the greatest hopes are pinned on further work on the Herculaneum scrolls. Scientists estimate that there could be up to 10,000 scrolls in the villa – most still waiting to be discovered. If even a portion of them can be read, we can expect:
- New works by Greek and Roman authors whose texts were lost.
- Detailed information about daily life in ancient Rome (letters, receipts, notes).
- A better understanding of the relationship between Greek and Roman culture.
The future of reading damaged texts: What awaits us?
The Vesuvius Challenge is just the beginning. In the coming years, we can expect:
- Improved algorithms: Scientists are working on models that handle scroll deformations and low-resolution scans better.
- Faster scanners: New X-ray tomographs will allow for scanning more scrolls in less time.
- Expansion to other languages: Currently, algorithms are trained mainly on Greek. In the future, they may include Latin, Hebrew, and even undeciphered languages like Etruscan.
One of the most exciting directions is combining AI with robotics. Researchers at the University of Kentucky are working on a system that could automatically scan and analyze scrolls without human intervention. This would significantly speed up work, although it also raises ethical questions – should we "open" all scrolls, even those that might contain sensitive content?
Regardless of the answer, one thing is certain: artificial intelligence has forever changed the way we study the past. As Nat Friedman, one of the founders of the Vesuvius Challenge, notes: "This is not just about reading a few ancient texts. It's about opening the door to a new era in archaeology, where machines help us recover the voices of those who have been silent for millennia."
Are we facing an era where AI becomes a standard tool for historians? Probably yes – but like any tool, it will require responsible use. After all, reading the text is only the first step. The real challenge begins when we try to understand what it means for us.
Further reading
If the topic of artificial intelligence in science has interested you, we recommend other posts on our blog:
- Autonomous AI as a new species: Why are we heading toward the total replacement of humans? – on the future of artificial intelligence and its impact on society.
- Local AI: The Odysseus Project as an alternative to centralized platforms – on distributed machine learning systems.
- Ubuntu 26.04 LTS as an operating system for AI and machine learning – a technological analysis of tools for researchers.
Sources
- https://geekweek.interia.pl/historia/news-filozof-sprzed-2200-lat-przemowil-ai-odczytala-zweglony-zwoj,nId,23505015
- http://www.herculaneum.ox.ac.uk/
- https://www.britannica.com/topic/Herculaneum-papyri
- https://scrollprize.org/
- https://www.cs.uky.edu/dri/
- https://www.nature.com/articles/s41467-019-11575-0
- https://scrollprize.org/firstletters
- https://github.com/ScrollPrize
- https://www.deadseascrolls.org.il/
- https://sinaipalimpsests.org/
- https://www.nature.com/articles/s41598-021-97422-6
- https://beinecke.library.yale.edu/collections/highlights/voynich-manuscript
- https://www.youtube.com/watch?v=example
Comments