> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# An AI model for restoring Ancient Greek papyri matched or beat human restorations in 77% of blind comparisons
- URL: https://www.metatalks.ai/ai-model-restoring-ancient-greek-papyri-matched-human-restorations/
- Published: 2026-09-29T17:43:00.000Z
- Updated: 2026-09-29T17:42:59.000Z
- Author: Al
- Tags: News, Science AI, #newswire

**Apollo, built by the Austrian Academy of Sciences with Mistral, also improved the reading of a Vesuvius-charred papyrus roll edited after its training data was compiled.**

The Austrian Academy of Sciences released Apollo in September as a free, open-access tool trained on about 600 million words of historical Greek. The French AI company Mistral helped build it on top of its Mistral Small model. Given a torn or damaged text, Apollo proposes several candidate words and passages for each gap, and a scholar picks among them, a design meant to keep the model's errors out of the historical record.

In a blind study described in the team's [paper](https://arxiv.org/abs/2609.22455?ref=metatalks.ai), 20 specialists in papyri, inscriptions and ancient texts compared Apollo's suggestions for genuinely unknown gaps with human restorations. They judged the model at least as good in 77% of cases and better in 16–20%. The paper also reports that Apollo improved the published reading of a roll carbonized when Vesuvius erupted in 79 CE, which scholars digitally unrolled and edited after the model's training data was compiled.

More than a million papyrus fragments survive in museums and libraries, and 92% still await deciphering, according to the academy. Restoring one takes a scholar who can split unspaced Greek into words, date the document and weigh its context before choosing words for the gaps. Only a handful of people anywhere are that good, Stephen Colvin, a classicist at University College London, told [Wired](https://www.wired.com/story/apollo-ai-model-ancient-greek-secrets-papyrus/?ref=metatalks.ai).

Apollo adapts to what it reads, says project lead Anna Dolganov, a papyrologist at the academy. Given Homer, it fills in Homeric Greek; given a Doric inscription, it answers in Doric. Its reconstruction of a fragmentary inscription from the Black Sea region turned up evidence of Rome's tax on prostitution, showing that Roman law already applied in that area.

Many papyri have stayed unrestored because they are mundane: letters, marriage contracts, civil-service paperwork. Colvin does not expect Apollo to turn up new plays by Sophocles. Dolganov says a new picture of ancient social and economic history is emerging with its help from documents in bulk, such as tax receipts.