Lost in Ambiguity: AI and the Limits of Processing Turkish Morphology

Abstract

Despite rapid progress in natural language processing, artificial intelligence (AI) systems continue to struggle with languages characterized by high morphological complexity, such as Turkish. A central difficulty lies in their limited ability to resolve ambiguity, which is deeply embedded in Turkish through features like flexible word order, extensive case marking, and agglutinative structures. Current AI models, primarily shaped by exposure to massive corpora of performance data, excel at detecting surface-level statistical regularities but fall short of grasping the grammatical principles and semantic nuances that underpin genuine linguistic competence. This reliance on performance over competence makes disambiguation especially problematic, as the systems lack the deep structural awareness required to interpret sentences where multiple readings are possible. To explore these weaknesses, the study tested five prominent AI models, Gemini, Claude, ChatGPT, Deepseek, and Grok, using ten carefully selected ambiguous Turkish sentences. Without being alerted to the ambiguity, the systems frequently generated interpretations that were semantically inappropriate, inconsistent, or distorted by hallucinated content. The findings illustrate how data-driven training alone cannot equip AI with the pragmatic reasoning and world knowledge necessary for accurate interpretation. The paper argues for a shift toward models that integrate richer linguistic theory, enabling AI to move beyond statistical mimicry toward a more human-like capacity for language understanding. Such an approach is vital for developing tools that can handle morphologically rich and ambiguity-prone languages with greater fidelity.