Effective vocabulary acquisition in foreign language learning extends beyond memorization to understanding nuanced distinctions between synonyms and their contextually appropriate uses. Traditional vocabulary instruction methods often fail to address these complexities, particularly for learners who struggle with subtle semantic differences. This study developed and validated the Contextual Adaptive Vocabulary Enhancement System (CAVES), an AI-driven platform designed to enhance vocabulary learning through personalized, context-rich exercises. CAVES integrates OpenAI’s GPT-4 to generate fill-in-the-blank questions with varying ambiguity levels and employs BERT-based models for probabilistic evaluation and adaptive feedback. The system supports multiple languages and enables learners to customize their practice sessions by selecting target synonyms, Common European Framework of Reference for Languages proficiency levels, and question quantities. A comprehensive experimental validation involving 88 (46 Japanese English learners and 42 native English speakers) participants was conducted. They completed 20 automatically generated English preposition questions. Statistical analyses revealed exceptional reliability (Cronbach’s α = 0.88) and a strong negative correlation between the AI-assigned ambiguity levels and participant performance (r = −0.76, p < 0.001). Native speakers significantly outperformed Japanese learners (M = 0.92 vs. M = 0.68, p < 0.001), confirming the sensitivity of the system to proficiency differences. These results demonstrate CAVES’ effectiveness in generating appropriately challenging vocabulary assessments and providing meaningful differentiation between learner proficiency levels, thus establishing its value for computer-assisted language learning applications.