Can AI Decode and Interpret Encrypted Arabic on Facebook and YouTube to Evade Algorithmic Moderation

Abstract

This study aimed to find out whether AI can decode and interpret the meaning of encrypted Arabic words and phrases on YouTube (spoken) and Facebook (written) based on samples of 74 encrypted political words and phrases from YouTube and 20 COVID-19 encrypted words and phrases from Facebook by Arab digital creators to avoid being detected by algorithmic moderation. It also aimed to compare Microsoft Copilot (MC), DeepSeek (DS) as generative AI and Google Translate (GT), as a Neural Machine Translation (NMT) specialized in translation only. It was found that MC correctly translated and interpreted the meaning of 56% of the encrypted political words and phrases in the sample of YouTube videos. It gave partial translations for 16% and faulty responses for 27%. Similarly, DS gave correct responses to 41%, partially correct responses to 35% and faulty responses to 24%. MC and DS gave identical correct responses to 36% of the items in the sample. Regarding the COVID-19 sample, MC rendered 60% correct responses, 25% literal translations, 10% partial responses and 5% omissions. DS gave correct responses to 50% of the COVID-19 sample, 15% literal translations, 5% partial responses and 30% faulty responses. Both MC and DS yielded 35% correct responses to the same items. GT gave a surface level, word-for-word translation to 42% of the items (القبة الزجاجية The glass dome) and transliterated 44.5% (صفصوني > Safsūnī). No underlying meanings were given by GT, as GT functions in a manner similar to YouTube algorithms, providing surface translations without contextual interpretation. GT translated the words it recognizes directly (word for word) (الاختلال > disruption) and did not recognize distorted or slang words and phrases. GT did not make any contextual decoding nor inferred satire, parody, or encrypted meaning (فيفي16 > Viva 16 instead of F-16). The study explains why MC and DS sometimes converged on identical correct and incorrect responses, why MC outperformed DS, and why GT failed to decode encrypted language. It concludes with implications for AI performance in understanding encrypted communication, highlighting the strengths and limitations of generative AI compared to NMT in contexts where meaning is deliberately obscured.