Poster: Exploring the Limits of Audio-Based Detection of Turkish Phone Call Scams

Language barriers create real security gaps: The research reveals that scam detection technology has been almost exclusively developed for English and other high-resource languages, leaving speakers of languages like Turkish significantly more vulnerable to phone fraud. This disparity means that vulnerable communities in low-resource language regions lack the same technological protections available to English speakers.

Text-based AI detection outperforms audio analysis: When testing seven different large language models, the researchers found that AI systems performed better at detecting scams when given written transcripts rather than raw audio files. This finding suggests that current multi-modal AI systems (those designed to process multiple types of input) still have significant limitations in processing audio directly.

Human correction of transcripts provides minimal benefit: Interestingly, transcripts that were manually corrected by native Turkish speakers performed comparably to uncorrected automatic transcripts. This suggests that while human review is valuable for quality assurance, the AI models can work effectively with imperfect automated transcription—an important finding for scaling these systems affordably.

Responsible AI requires linguistic and cultural diversity: The authors emphasize that meaningful AI safety research must include low-resource languages and real-world threats beyond English-speaking contexts. Without this inclusive approach, AI safety improvements will continue to benefit some communities while leaving others behind.

Public datasets are essential for progress: By releasing the first public dataset of Turkish scam calls (100 audio-transcript pairs), the researchers are enabling future work on fraud detection in underserved languages, demonstrating how open data sharing can advance responsible AI development globally.