From words to pixels: Artificial intelligence struggles with world Englishes

Abstract

This study examines how DALL-E 3 interprets English descriptions written by Indonesian university students. Sixteen descriptive texts were submitted to the artificial intelligence (AI) tool, and the resulting images were compared to original photos. Most outputs showed clear mismatches. The analysis found that misinterpretations originated from two main sources: grammatical and vocabulary patterns reflecting Indonesian English and broader stylistic choices, such as the use of vague, emotional, or abstract language. The study also found that a high level of concrete detail could often mitigate the negative effects of non-standard grammar. The findings suggest that current AI tools are not yet equipped to fairly process the full range of human linguistic variation, from local English features to the stylistic patterns of human-centric writing. To support more inclusive use of AI in education, this study adapts the established concept of intelligibility into the idea of “digital intelligibility,” and recommends improving training data and creating classroom space for open discussions about AI bias toward language diversity and stylistic choices.