LEXICAL CREATIVITY OF LARGE LANGUAGE MODELS: EVIDENCE FROM SUFFIXATION

Abstract

In this study, we examine lexical creativity in Large Language Models (LLMs) through a focused analysis of suffixation, a highly productive and well-described word-formation process in English. Using a selection of suffixation outputs generated by 24 LLMs (Dinu et al. 2025a), the study operationalises lexical creativity through a transparent, theory-driven scoring framework based on morphological well-formedness, productivity and semantic interpretability. Rather than treating creativity as a psychological construct or attempting to measure it directly, our analysis follows a descriptive and exploratory approach, relying on linguistically motivated composite indices. The results show that LLMs demonstrate strong morphological competence in suffixation, while creative behaviour varies across the two suffixes (-ish vs. -able) and across the tested LLMs. Creativity emerges most clearly at the boundaries of productivity, where increased novelty often goes hand in hand with overgeneralisation. The findings suggest that lexical creativity in LLMs involves exploring the limits of existing word-formation patterns, rather than creating words without constraints, and highlights how artificial language systems manage the balance between structure and innovation.