Using AI to complete missing portions of Roman inscription

[guy]

 

Researchers are now using a neurogenerative AI, named Aeneas, to fill in missing parts of Roman inscriptions. The military bronze diploma from Sardinia shown above, for example, was issued by Emperor Trajan to a sailor. Using Aeneas, researchers believe they were able to restore missing sections of the text.

This study introduces Aeneas, a multimodal, generative neural network designed to assist historians in analyzing and interpreting ancient Latin inscriptions. It addresses three key challenges in epigraphy: text restoration, geographical attribution, and dating chronology.

🔍 Key Features of Aeneas

Contextualization: Finds historically relevant parallels to inscriptions, aiding interpretation.

Multimodal Input: Uses both text and images of inscriptions.

Arbitrary-Length Restoration: Can restore missing text even when the length is unknown.

High Accuracy: Dates inscriptions within ~13 years and correctly identifies the origin of the inscription in the Roman provinces 72% of the time.

🧪 Evaluation Highlights

In a large-scale study with 23 historians:

◦ Aeneas’ parallels were useful in 90% of cases.

◦ Historians’ confidence increased by 44% when using Aeneas.

◦ Combined human–AI performance outperformed each alone.

🏛️ Case Studies

• Res Gestae Divi Augusti: Aeneas accurately identified linguistic and historical markers to date and contextualize this imperial inscription.
• Votive Altar from Mainz: Retrieved a rare parallel inscription, demonstrating its ability to detect subtle historical connections.

 

https://www.nature.com/articles/s41586-025-09292-5

 

https://deepmind.google/discover/blog/aeneas-transforms-how-historians-connect-the-past/

 

 

Edited July 27, 2025 by guy

[caesar novus]

I've recently seen some amazing transcripts (on X)  of folks asking history questions to AI machines. I mean in a skilled probe that avoids regurgitating unsupported claims. Like examining billionaires who ultimately were based on crime. I hear the next version of chatgpt will integrate multimedia to handle images, video, text all together as needed vs the elaborate way you have to tailor "query engineering" now.

A nitpick is the video here seems to be the AI generated (chatgpt?) dialog mode which appears so phony, limited, and contrived to me. They can and should flip a switch and have it come out as a straightforward narration or exposition instead. Grok will next come out with auto multiple AI agents in parallel that actually could give interesting dialogs to debate variations of answers.