We often talk about AI’s impact on modern communication, but its role in understanding the ancient world is becoming just as compelling. A new study in Nature shows how advanced neural networks can support historians in interpreting damaged inscriptions, offering insights that were once locked behind missing text, eroded surfaces & fragmented context.
The study
Researchers introduced Aeneas, a multimodal generative neural network trained on nearly 177,000 Latin inscriptions, dating from the 7th century BCE to the 8th century CE. The dataset included both transcriptions & images where available.
Aeneas was tested on three notoriously complex tasks in epigraphy:
- Restoration of missing text (including sequences where the number of missing characters is unknown, known as lacunae).
- Geographical attribution (predicting the Roman province of origin).
- Chronological attribution (dating the inscription by decade)
Crucially, Aeneas doesn’t just “guess”. It generates contextual parallels—related inscriptions with similar linguistic, stylistic or cultural features—to show its working. It operates like an eager academic assistant offering suggestions: “Here are ten similar examples from across the empire. Now you decide.”
To gauge usefulness, the researchers ran the largest historian–AI collaboration of its kind: 23 expert historians evaluated Aeneas’ suggestions across controlled tasks.
The findings
Aeneas significantly boosted historians’ performance & confidence:
- Useful starting points: Historians judged Aeneas’ parallels helpful in 75% of cases, rising to 90% when predictions were also provided.
- Confidence: Increased by 23% with parallels & 44% when predictions were included.
- Restoration: Historians alone had a character error rate of 39%; with Aeneas’ predictions this fell to 21%.
- Geographical attribution: Humans alone hit 27% top-1 accuracy; with Aeneas this reached 68%, outperforming either humans or AI independently.
- Dating: Human predictions averaged 31 years from the true date; with Aeneas this dropped to 14.1 years. Aeneas alone averaged 12.8 years.
In other words, the partnership of human expertise & machine pattern-spotting consistently outperformed either working solo.
Relevant context
This work sits within a growing body of research using machine learning for ancient languages. It builds on earlier models such as Ithaca (Assael et al., 2022), but extends capabilities:
- handles unknown-length lacunae- (hugely important for real inscriptions)
- integrates visual features (e.g. letter shapes, layout, material)
- retrieves historically meaningful parallels, not just string matches
From a linguistic perspective, Aeneas illustrates how deeply distributional learning can capture diachronic change (ie over time). For example, the model picks up on spelling shifts (“aheneis” > “aeneis”), political titles linked to specific historical periods, or culturally marked personal names—just as a trained historian would.
An ELT equivalent: imagine trying to place an English text in time using only linguistic clues. The spelling olde, the phrase thou art, or references to “the wireless” would signal different eras. Aeneas does this kind of triangulation across 2,000 years of Latin.
Teacher Takeaways?
No direct implications for ELT-practice, but this makes me think that:
Even highly specialised experts benefit from AI-curated input, not because AI “knows more”, but because it accelerates the search for relevant patterns. This mirrors language classrooms: curated exposure often matters more than sheer quantity.
The study highlights how examples drive inference. Whether restoring Latin or acquiring English, learners rely on parallels to build mental models.
Aeneas underscores the value of multimodal input. For language learners, combining text with images or context isn’t just decorative; it’s cognitively supportive.
How might tools like this shape the way we approach language, history & evidence in our own classrooms?
Leave a Reply