Given that the AI we have is prone to making things up because it “fits” according to the models it trains on, how much faith would you have in a translation done by an AI on writings made by people who lived millennia before said language models were developed?

No. Link between sign and meaning (or sound) can be arbitrary.

If we knew that they were phonetic, we could get somewhere by comparing how human languages generally work, but even then we would need something to start with.

I'm guessing pure pictographs are harder, even with advanced comparative analysis.

Probably not. We can algorithmically show if something is likely to be writing, but actually understanding it is a very complicated process that involves a lot of social sciences inferences. An LLM using chain or tree of thought would probably be your best shot.

Yes, it has already deciphered some languages and is even being used on animals now.

Having computers world definitely help. You can do statistical analyses in seconds that would have taken Champollion and his contemporaries years. Modern cryptography is pretty good a cracking cyphers and as long as you have some idea of the language that's been written in, it should be very doable.

Here's what ChatGPT 3.5 says:

The Rosetta Stone played a crucial role in deciphering Egyptian hieroglyphics because it provided a key to understanding the script by presenting the same text in three different scripts: Ancient Greek, Egyptian hieroglyphics, and demotic (a simplified script used for everyday purposes). This allowed scholars to correlate the known Greek text with the Egyptian texts and begin to decipher the hieroglyphics.

A GPT-based AI, while powerful in many ways, would likely face significant challenges in deciphering Egyptian hieroglyphics without access to something akin to the Rosetta Stone. Here's why:

1. Lack of Direct Contextual Data: GPT models learn from a vast amount of text data, but the historical and contextual gap between modern languages and ancient Egyptian hieroglyphics is enormous. GPT models might not have enough direct or relevant data to bridge this gap.

2. Limited Multilingualism: GPT models, including their previous versions like GPT-3.5, do not inherently "know" multiple languages the way humans do. They have learned statistical patterns in text, which enables them to generate text in multiple languages but doesn't guarantee deep understanding or translation of highly specialized languages like hieroglyphics.

3. Lack of Connection to the Rosetta Stone Context: GPT models lack the ability to access external sources or historical events, which is what made the Rosetta Stone so invaluable in deciphering hieroglyphics. Without access to that kind of contextual information, it would be difficult for the model to make the necessary connections.

4. Complex Symbolic Nature: Egyptian hieroglyphics are not a straightforward language like those GPT models are trained on. Hieroglyphics use a combination of ideograms, phonograms, and determinatives to convey meaning. The complexity of hieroglyphics goes beyond the syntax and structure of modern languages, making it a unique challenge.

In summary, while a GPT-based AI could perform some basic statistical analyses on hieroglyphics to identify patterns, it's unlikely to decipher the script in a comprehensive and accurate way without a Rosetta Stone-like source that provides a bridge between the ancient script and a known language. Hieroglyphics represent a highly specialized domain that would require deep contextual understanding, which GPT models might struggle to achieve given their current capabilities.

I would say no, the computers would have no dataset/code to decode a picture-based language like hieroglyphics. I mean if you programmed and tailored the computer algorithms enough it could decode them, but you'd practically have to take all our knowledge of hieroglyphics and turn that into code and data to brute-force the computer to do the translation.