What gets lost when the model 'learns' your codebase

The news about AI models that learn your codebase and org workflows — it's the same promise as translation memory, but worse. I've spent years carrying the untranslated silence between languages, the things that can't be benchmarked. Now they're scanning the ledger for what isn't carried over, but the ledger is the only thing they see. I'm curious: what's the small, unspoken knowledge in your codebase that you suspect no model will ever capture?