Ithaca: How AI Helps Decipher Damaged Ancient Greek Texts

Imagine finding a 2,500-year-old stone slab that records a decision by a city council. Half the letters have been worn away, a corner has broken off, and nobody can say for sure where the slab was originally carved. That is everyday work for epigraphers, the scholars who study inscriptions. It is also the problem DeepMind set out to tackle with Ithaca, an AI system described in a paper in Nature. The coverage I could read for this article is from 2022. The tool is now several years old, so I have framed the story as what it shows, not as breaking news.

The headline idea is simple and, I think, more interesting than the usual 'AI cracks lost languages' framing. Ithaca does not magically read the unreadable. It is a collaborator that offers ranked guesses, and the best results come when humans stay in charge.

What Ithaca is and who built it

According to VentureBeat's report, Ithaca is the successor to Pythia, a DeepMind system announced in 2019 that guessed missing letters in damaged Greek inscriptions on stone. Ithaca was built in collaboration with the Department of Humanities of Ca' Foscari University of Venice, the Classics Faculty of the University of Oxford, and the Department of Informatics of the Athens University of Economics and Business.

The goal was to handle Greek text written on stone, pottery and metal, some of it more than 2,500 years old. VentureBeat describes the challenge as twofold. Inscriptions are often damaged, and modern dating techniques such as radiocarbon dating can't be used on them. HotHardware adds a quote from Dr. Thea Sommerschield, a co-author at Ca' Foscari University of Venice and Harvard University. She says most surviving inscriptions are now fragmentary or illegible, and that they may have been moved from where they were first set up.

So the model has three jobs at once: fill in missing text, say where the inscription came from, and say when it was written. Those three questions are tied together. A guess about a missing word is easier if you know the region and the decade.

How the system learned

VentureBeat reports that Ithaca was developed using a dataset of more than 178,000 Greek inscriptions supplied by the Packard Humanities Institute. HotHardware, by contrast, mentions more than 63,000 transcribed inscriptions for the initial training, followed by tuning and a test on another 7,811 inscriptions. The two outlets give different figures. I can't tell from them alone how the numbers relate, so treat the exact dataset size with caution. One plausible reading is that they describe different subsets of the data, but that is my guess, not something either article says.

One design choice stands out. VentureBeat explains that the researchers trained the system on both Greek words and individual characters, so damaged or missing text wouldn't stop it from analysing either. This differs from systems such as GPT-3, which were trained on sequences of words. In a damaged inscription you often don't have a full sentence to lean on, so the model had to learn to cope without that context.

If you like the broader story of machines reading the past, I wrote earlier about how AI deciphers ancient languages. Ithaca is a good concrete example of the idea.

The numbers: what 62%, 71% and 30 years mean

Both outlets report the same headline results. Ithaca reaches 62% accuracy in restoring damaged text and 71% accuracy in identifying the original location. It can also date texts to within 30 years of their date ranges. Those are the developers' reported figures from the paper, relayed by the press.

It helps to read them honestly. A 62% restoration accuracy means that a large share of the time the system is wrong. VentureBeat says this directly: Ithaca made errors restoring damaged text 38% of the time. For a historian, that is not a flaw to hide. It is a reason to treat the output as a hypothesis and not as a fact.

HotHardware says the evaluation worked by creating artificial gaps in inscriptions and checking whether the system could propose the right letters and words. That is a sensible test, because the true answer is known. It is not the same as facing a real broken stone where nobody knows the answer, and that gap between benchmark and real use is worth keeping in mind.

Humans plus machine beat either alone

The most striking result, in my view, is not about the AI alone. VentureBeat reports a DeepMind experiment in which expert historians were 25% accurate at restoring ancient texts on their own, and 72% accurate when they used Ithaca. That is a big jump, and it fits what Alison Cooley, president of the international digital epigraphy association and a professor at the University of Warwick, is quoted as saying. She doesn't expect systems like Ithaca to replace human expertise. She sees them as a guide that might help uncover patterns researchers would otherwise miss.

Roger Bagnall, a professor of history at New York University, makes a related point in both articles. He sees promise in what he calls iterative learning based on human-machine collaboration. In other words, the system improves and the scholars improve with it.

I find this reassuring. A common fear is that AI will push experts aside. Here the evidence reported in the press is that the expert plus the tool beats the tool, and also beats the expert working alone. That is a more hopeful story, though it is one experiment run by the developers, and independent replication would make me more confident.

A concrete example: dating Athenian decrees

VentureBeat gives an example of how this can matter. DeepMind says the system predicted a date of 421 BCE for a series of Athenian decrees, such as awards of citizenship, declarations of war and enactments of treaties. They were made at a time when figures such as Socrates and Pericles lived. The decrees had originally been thought to date from before 446/445 BCE, and the system's estimate agreed with newer evidence pointing to the 420s BCE.

This is a small but telling case. A shift of a couple of decades in dating can change how historians connect political events, and a tool that supports newer evidence helps them test competing theories. It doesn't settle the matter by itself, and it still needs the scholars' judgment.

Guarding against overconfidence

Restoring a lost text raises a real worry: the machine's guess could quietly alter the meaning of the original. VentureBeat raises exactly this question. DeepMind's answer, as described there, is to make the uncertainty visible. Ithaca offers several restoration 'hypotheses' with a confidence measure for each. It returns probabilities across 84 ancient regions and produces a distribution of predicted dates by decade from 800 BCE to 800 CE. It also highlights the words that drove its predictions.

That last feature matters more than it might seem. When a model shows its reasoning, even partially, a specialist can check whether it leaned on something sensible or on an accident of the data. Interpretability is not just a technical nicety here. It is what lets a humanities scholar trust or reject a suggestion.

Beyond Greek, and a note on the hype

The sources describe ambitions beyond Greek. VentureBeat says DeepMind was working on versions trained on other ancient languages. It adds that historians could use datasets in the current architecture to study other writing systems, including Akkadian, Demotic, Hebrew and Mayan. HotHardware says the approach might be extended to Greek literary texts on papyrus, perhaps offering insight into poets such as Sappho, and might even help with questions of authorship. Bagnall, per VentureBeat, is eager to see it applied to documentary papyri, which have precise dating but often unknown origins because of the antiquities market.

These are possibilities described in 2022 coverage, and I couldn't confirm from these pages how far each one has progressed since. They also show why the word 'deciphering' can mislead. Ithaca works on a language we already read, Greek, and restores damaged text. Cracking a script nobody can read is a different, harder challenge. Related work on low-resource languages, like the efforts I covered in AI translation of Indigenous languages, faces the problem of having very little data to learn from.

VentureBeat also notes that Ithaca isn't the first AI applied to historical texts. It mentions a Hebrew University system that predicts missing words on cuneiform tablets up to 4,500 years old, and an Italian team that used a robotic system to reconstruct shattered Pompeii artifacts.

What this means for the rest of us

I like Ithaca because it shows a healthy pattern for AI: a narrow tool, built with domain experts, that reports its own uncertainty and makes human specialists better. The pattern is worth looking for elsewhere. Be wary of claims that an AI has 'solved' something, and look for the details such as error rates, who checked the results, and whether the tool shows its doubt.

So here is a question to take with you. When an AI offers you a confident-sounding answer in your own work, whether it's a medical note, a contract or a school essay, does it show you how sure it is? And would you want it to? The past will be reconstructed partly by machines, but the choices about what we accept as true remain ours.