Every major AI model on market today fails at polytonic Ancient Greek. Not a little. Completely. Ask ChatGPT to parse a sentence from Aristotle's Nicomachean Ethics in original Greek and watch what happens. It confuses accents, drops breathings, produces Modern Greek where polytonic should be. The entire Corpus Aristotelicum, 2,400 years of philosophical reasoning, is invisible to the systems we call "intelligent."
The reason is technical but simple. Training data for Ancient Greek is almost zero. Modern Greek exists in some quantity, but polytonic script, with its rough and smooth breathings, acute, grave, and circumflex accents, is a different orthographic system entirely. RLHF training made problem worse. Human raters do not know polytonic Greek. They rate outputs based on what looks reasonable to them, which means they reward Modern Greek approximations and punish authentic polytonic forms. The alignment process systematically destroys what little Ancient Greek capability the base model had.
This is not just philology problem. It is architecture problem. Every time you fine-tune for "helpfulness" and "safety" you compress reasoning space. The model becomes better at producing plausible-sounding English and worse at everything else. Polytonic Greek is first casualty because training signal for it is weakest. But same mechanism affects any domain where authentic reasoning diverges from what average rater considers helpful.
Classics departments in universities are shrinking. Enrollment drops every year. Fewer students learn Ancient Greek. Fewer professors can teach it. And now AI tools that could help preserve and study these texts are actively degraded by training processes designed to make them "safe." The irony is complete. We build systems that cannot read what we most need them to read.
Some labs tried to fix this with more data. Bigger pre-training corpus. But problem is not pre-training. Base model Qwen or Llama can actually handle polytonic Greek at low level, it recognizes characters, produces diacritics. RLHF training on top destroys this capability because reward model has no signal for correctness in Ancient Greek. You cannot align what you cannot evaluate.
Solution is not bigger models. It is different architecture. Corpus-grounded systems that retrieve from source texts in original language, that do not rely on parametric knowledge alone. RAG over digitized critical editions. Systems trained to respect form of source material rather than rewrite it into what rater expects.
We work on this. Not with ChatGPT API but with self-hosted models, RAG pipelines over digitized Greek texts, evaluation metrics that actually check polytonic accuracy. It is hard. Very hard. But somebody must do it because alternative is losing entire tradition of Western philosophy to training-data bias.
The word for this is παιδεία. Not just education but cultivation of intellectual capacity through engagement with difficult texts. Your AI cannot provide παιδεία because it was trained to avoid difficulty. It gives you summary when you need source, paraphrase when you need argument, Modern Greek when you need polytonic.
Corpus problem is not edge case. It is proof that alignment as currently practiced destroys knowledge. Not by accident. By design. When you optimize for average rater satisfaction you lose everything that falls outside average rater competence. And most important knowledge in human history falls exactly there.
What happens to discipline when its primary texts become unreadable to primary research tools of next generation?
[link] [comments]