Why Generic AI Translation Loses Narrative Continuity
A chapter that ends "she finally answered him" should translate so the next chapter's "he turned away" still points at the same two people. In chat-based AI translation, it often does not. The pronoun lands on the wrong character, the tone shifts from intimate to neutral, the unresolved question from a page earlier becomes a fresh statement. The model is not at fault. There is simply nothing carrying the story from one request to the next.
What goes wrong
Pronouns flip gender when the target language requires marking. A speaker introduced two paragraphs ago is treated as a stranger. A formal register established in chapter one collapses into the model's default by chapter three. The technical term agreed in the methods section is rephrased in the discussion.
Why generic AI translation fails here
Every chunk is a fresh HTTP call. The model has no memory of the previous chunk. Pasting the whole chapter as context fills the window and inflates cost. Conversation history works for a few turns and then the provider drops the early turns silently. There is no place for context to live that survives a model swap or a retry.
How TranslationAI solves it
The tail of each chunk is stored as context for the next chunk, in the database, marked so the model uses it for understanding but does not translate it. Because context lives outside the model, it survives provider fallback and retries. A chunk that fails on one provider and retries on another carries the same context, so the continuity does not reset.
Further reading: three root causes of consistency drift, why AI fails on book-length documents, register and why it must be preserved.