This study examines the role of Human-Centered AI (HCAI) in AI-assisted software localization by assessing ChatGPT as a localization tool for culturally sensitive Arabic app adaptation. Using an open-source mobile media application as a case study, the study evaluates how large language models handle the transformation of domain-specific interface terminology into culturally appropriate Arabic equivalents across 213 translatable XML strings. The experiment was conducted longitudinally by comparing two model generations, ChatGPT-4 and ChatGPT-5.4, using the same source file, prompting sequence, and evaluation framework. Quality was assessed through cosmetic and functional testing, alongside linguistic analysis based on a purpose-built eight-category error taxonomy extending existing localization quality assessment models. ChatGPT-4 produced errors in 49.8% of strings, compared with 45.5% for ChatGPT-5.4. Although the overall error-rate difference was not statistically significant, the distribution of error types shifted substantially. ChatGPT-5.4 showed improvement in semantic completeness and terminological consistency, but introduced more literal UI mappings and continued to struggle with ontological transfer, particularly in distinguishing domain-neutral technical concepts from culturally marked domain concepts. The findings demonstrate that LLMs can support localization workflows as first-pass generators, but they cannot replace expert human judgment in culturally sensitive domains. Human expertise remains essential for prompt design, terminology control, cultural calibration, ontological mapping, and final quality assurance. The study therefore provides empirical support for the HCAI paradigm in translation and localization, showing that effective AI use depends not on automation alone, but on expert-led human–AI collaboration. It also highlights the need for longitudinal evaluation of LLMs, as newer model versions may improve some error categories while introducing new risks.