LLM translations add emotional fingerprints, altering author's voice in literature
AI translators leave model-specific emotional traces even after human editing.
Deep Dive
A study compared LLM translations of Margaret Atwood's *Oryx and Crake* with post-edited versions and a human translation. Using lexicon-based and multilingual emotion modeling, the study found that each machine translation (MT) system introduces statistically significant, model-specific emotional fingerprints. Even post-editing didn't fully erase these shifts, resulting in limited preservation of the author's original voice.
Key Points
- Each LLM translation model (e.g., GPT, Claude) introduces statistically distinct emotional fingerprints compared to human translators.
- Post-editing reduces but does not eliminate these systematic emotional shifts from the original text's authorial voice.
- Study used both lexicon-based and multilingual emotion modeling on a large Italian sci-fi corpus as baseline for comparison.
Why It Matters
Literary translators and publishers must account for AI's emotional bias to preserve an author's voice in translated works.