Fabricator or dynamic translator?

The paper's core contribution is a taxonomy, not just a detection system: it argues that not all spurious text in machine translation is harmful, and that conflating helpful clarifications with hallucinations causes commercial systems to over-suppress useful output. That distinction is the buried lede the summary softens.

This connects directly to the reliability measurement problem surfaced in 'Diagnosing LLM Judge Reliability' (also from April 16 on arXiv), which found that aggregate consistency scores mask per-instance logical failures in LLM evaluation. The same structural issue applies here: if detection systems in commercial MT pipelines are trained on coarse labels that lump clarifications with hallucinations, they will produce high aggregate precision while systematically misclassifying a meaningful subset of outputs. Both papers are, at root, about the gap between summary-level metrics and instance-level correctness. The DiscoTrace work from the same period adds a related angle, showing that LLMs already favor breadth over selectivity in how they construct responses, which may partly explain why spurious elaboration appears so frequently in translation contexts.

Watch whether any of the major commercial MT providers (DeepL, Google Translate, ModernMT) publish updated quality documentation that distinguishes hallucination rates from clarification insertion rates. If that split appears in product benchmarks within the next two quarters, this taxonomy is gaining traction outside academia.