Decoding Digital Evasion: Recognition and Interpretation of Obfuscated Arabic Text in Images by Gemini, Copilot, Claude, and Qwen

Abstract

This study examines how four AI models—Gemini, Copilot, Claude, and Qwen 3.7 Plus—decode and interpret obfuscated Arabic words embedded in 138 images extracted from YouTube and Instagram videos posted during the USA–Israel–Iran War (March–April 2026). The obfuscation relied primarily on inserting punctuation marks (periods, commas, dashes, underscores) to evade platform moderation. Across these images, the dataset contained 374 obfuscated Arabic words, which the models translated in context and in isolation. Analysis revealed variations in recognition and interpretation across the models. All four models correctly translated 30% of images; three models succeeded in 24%; and two models converged on 27%. Gemini produced uniquely correct translations for 16% of the images. All four models failed completely on 3% of the images, where root semantics were unrecoverable. Additionally, Gemini translated Arabic text in 94% of the images correctly, producing accurate translations for 98% of the words in context and 99% in isolation. Copilot succeeded on 61% of the images, correctly translating 72% in context and 98% in isolation. Qwen translated 54% of the images, yielding correct translations for 72% of the words in context and 98% in isolation. Claude also succeeded on 54% of the images, producing 67% correct words in context and 98% in isolation. Across all models, recognition accuracy was consistently higher on isolated words than within full images, indicating that failures were driven by visual noise, contextual interference, and safety‑filter activation, rather than linguistic limitations. Additional behaviors included deleted and replaced tokens, hallucinated translations, transliteration, and partial clipping. Copilot blocked few images or produced gibberish output. Gemini sometimes mirrored obfuscation patterns in English equivalents. In isolated‑word translation, the models produced identical and synonymous equivalents. Claude added feminine gender markers to some distorted words.  The lexemes exhibiting the highest number of polymorphic distortion variants were Iran, Israel, war, missiles, nuclear, America, attack, Netanyahu, party/war, Gaza and Zionist. How LLMs process obfuscated Arabic text, density, complexity, and symbol typology of obfuscation, contextual vs. isolated processing, the spectrum of mimetic obfuscation, non-obfuscation hallucination, algorithmic paranoia, keyword panic, and socio-cultural misalignment are discussed in detail.