AI translation workflow

Why ChatGPT web novel translation drifts across chapters

ChatGPT can translate a scene well. The harder problem starts when the same story continues for dozens of chapters. Names, character voice, recurring terms, and prior localization choices begin to drift unless the workflow keeps them visible.

One chapter is a translation task. A series is a memory task.

A normal chat window starts each request with limited context. Even when the model is strong, it does not automatically know which title translation you approved in episode 1, which insult should stay rough, which character speaks in mock-politeness, or which sound effect style you chose earlier.

That is why long-form AI translation often feels accurate in isolation but inconsistent as a book. The plot may survive, while the texture slowly changes.

The most common drift problems

• Names and titles receive different translations in later chapters.
• Fantasy roles become literal or awkward, such as "subjugator" when "slayer" or "hunter" reads better.
• Rough dialogue becomes too polite, flattening character personality.
• Comedy timing gets smoothed into standard prose.
• Sound effects are translated inconsistently.
• Earlier reveals or relationship details are forgotten during later scenes.

Project memory is the practical fix

The fix is not a single perfect prompt. A prompt helps, but serial fiction needs reusable project context. A good workflow keeps approved decisions close to the next translation request: prior chapter summaries, saved translations, glossary entries, character notes, and style examples from edited output.

LoreLingo is built around that idea. Instead of treating every chapter as a blank text box, it stores the project decisions that should carry forward.

What to save before translating the next chapter

1. Approved name and title translations.
2. World terms, skill names, places, ranks, factions, and items.
3. Character speech rules: formal, rude, playful, archaic, blunt, nervous, or sarcastic.
4. Prior edited translation samples that show the intended target-language rhythm.
5. Short memory notes about important events and relationships.

Use review passes, not blind trust

Even with memory, AI output should be reviewed. A focused review pass should check fidelity, localization, terminology, character voice, and tone drift. This is especially important if the translation will be published, sold, or sent to readers.

The goal is not to make AI translation invisible. The goal is to make the workflow repeatable enough that each chapter starts from the same translation decisions.

LoreLingo is a Windows desktop app for this workflow: Korean, Japanese, and English web novel translation with project memory, glossary rules, character notes, batch translation, saved episodes, and review reports.