It isn’t quite as crazy as it sounds when you consider that a lot of inscription texts are pretty formulaic—epitaphs, dedications, and such. Plus, we have plenty of surviving writings in classical Latin, so we know the grammar pretty well. Given those things, I’d expect an AI trained on the corpus of inscription texts that have survived without significant damage to be able to make reasonable suggestions about formulaic texts.

Really, when you think about it, a trained human presented with a damaged inscription text won’t be doing anything much different from what an LLM would do: they’ll try to fill in the text with the most likely words based on any remaining traces of letters, and their knowledge of other, similar texts. The problem is getting the LLM to communicate its level of certainty about the fill-ins it’s offering.