← Azimuth Daily
Capturing data…
Capturing data…
SUN, 25 OCT · 77 ITEMS
"Commercial text-safety API misses harm and flags harmless text in Ukrainian and Surzhyk" — arXiv cs.CL · AI & Technology
This study evaluates a commercial (likely OpenAI's moderation system) for its accuracy in detecting harmful content across English, Ukrainian, and (a mixed Ukrainian-Russian language). The results show that the API misses harmful content and falsely flags harmless content at higher rates for Ukrainian and Surzhyk compared to English, revealing significant language bias in safety moderation tools.
The claim is based on a primary source: the arXiv preprint 'SITO: Sifting a Commercial Text-Safety API for Missed Harm and in English, Ukrainian and Surzhyk' by Andrii Tuzhykov. As a direct research paper, it provides detailed methodology and results, making it authoritative for its specific findings, though generalizability depends on the API and test set used.
The paper was published on arXiv on October 21, 2026, making it genuinely new research.