ArGuard draws 35 teams to Arabic AI safety

ArGuard’s Arabic AI-safety shared task drew 35 teams into its final evaluation from 58 registrants, with best macro-F1 scores ranging from 0.419 on fine-grained meme classification to 0.984 on one harmful-prompt track, according to the new benchmark paper. That spread matters more than headline participation: multimodal hate detection under distribution shift remained materially harder than prompt screening, so Arabic-language deployments need separate content and safety evaluations rather than one blended score. Teams building multilingual moderation can pair this result with the archive’s benchmark question audit before treating a single safety number as coverage.