Hallucinations in AI Translation: Structural Limitations Revealed by Odyssey and KBS Live Broadcast Errors

Recent controversies over AI translation errors have arisen following the release of classic translations tied to the box office success of the film 'Odyssey' and during KBS's live broadcast of the NASA Artemis II launch. In the AI-translated version of the Odyssey, Ancient Greek was mistranslated into modern slang such as 'Kingbatne' and 'Alppano?'. During the KBS live broadcast, technical air traffic control terminology 'Roger, roll, pitch' was displayed in subtitles containing profanity.

These incidents are not exceptional anomalies. In October 2025, a research team from Alibaba International Digital Commerce and Tianjin University evaluated 17 major LLMs, including GPT-4o-mini, Claude-3.7-Sonnet, and Gemini-2.0-Flash, using their proprietary benchmark 'HalloMTBench'. The results showed hallucination rates ranging from 33% to 60% depending on the model and language pair. Notably, models tuned with reasoning RL exhibited more severe errors, with hallucinations categorized into instruction deviation and source text deviation.

Structural Causes of Generative AI Translation

Generative AI operates by statistically generating the most plausible next token without a semantic representation separated from the source text. This approach reveals limitations where plausible expressions diverge from the source meaning, particularly in technical terminology and cultural contexts. The primary causes are as follows:

• Tokenization errors: Text is processed at the subword level, where agglutinative languages and neologisms are input as fragments rather than semantic units, potentially misaligning the foundational data from the start.
• Error accumulation in autoregressive structure: Since generated tokens become part of the next input, a single error is accepted as a premise without verification, accumulating errors until the end of the sentence.
• Misalignment of learning objectives: Models optimize for 'next token prediction accuracy' rather than directly targeting 'semantic equivalence'.
• Context dilution: In long contexts, the tone or metaphors from the beginning are not fully reflected at the time of generating the latter parts.
• Side effects of RLHF: Tuning processes aimed at producing natural output may displace subtle nuances of the source text.

Absence of Verification Layers and Non-Exposure of Confidence Signals

AI translation inherently lacks verification mechanisms. Unlike general software, generative AI does not treat misunderstandings of the source text as failures but instead outputs smooth, normal responses. Furthermore, although token probabilities (logprob) exist within the model, they are not exposed even in APIs such as Anthropic's Claude API, preventing users from distinguishing between low-confidence and high-confidence words. This demonstrates the structural limitation of AI translation operating without verification devices and confidence signals.