Bias in GenAI: Exploring cultural variations between large language models

• GenAI outputs exhibit recurring discursive tendencies of positivity in representations of culturally and linguistically responsive teaching (CLRT), with ChatGPT commonly construing CLRT through a depoliticised, “feel-good” multicultural frame, and Ernie Bot more often articulating harmony-centred representations. • While ChatGPT explicitly acknowledges the possibility of bias, its outputs tend to treat visual representations as self-evidently inclusive, limiting reflexive engagement with the representational constraints of images. By contrast, Ernie Bot adopts a more cautious stance, frequently foregrounding neutrality and balance, and at times downplaying bias through idealised portrayals. • Framing GenAI as a semiotic technology foregrounds how LLM outputs participate in the discursive construction and circulation of pedagogical imaginaries, rather than functioning as neutral reflections of pre-existing cultural ideologies. • These findings highlight the need for critical AI literacy that integrates CLRT with multiliteracies pedagogy, equipping educators and learners to interrogate how cultural diversity, language, and pedagogy are discursively framed and operationalised within increasingly AI-mediated educational environments. Ethical integration of generative AI (GenAI) in education demands urgent attention to addressing bias embedded in its content. Whilst considerable research has examined bias in GenAI models developed within Western contexts, less is known about how such bias presents itself in non-Western cultural contexts. This gap raises concerns about sociotechnical blindness where cultural assumptions embedded in GenAI models are unexamined and underscores the growing importance of culturally and linguistically responsive teaching (CLRT). This study examines how CLRT is interpreted and potentially misrepresented by two large language models, ChatGPT 4o and Ernie Bot 3.5 . By conceptualising GenAI as a semiotic technology, with large language models (LLMs) as a subtype, and adopting a multimodal critical discourse approach, we analysed 36 datasets comprising visual and textual outputs generated by the two LLMs in response to three standardised prompts focused on CLRT. Our analysis identified an inherent positivity bias across both LLMs which frame CLRT as depoliticised multiculturalism in ChatGPT and superficial harmony in Ernie Bot. Notably, ChatGPT recognises bias in relation to categorical exclusion and rights violation whereas Ernie Bot denies bias and frames it as absence of harmony or relational imbalance. We argue that LLMs do not merely reflect culturally entrenched ideologies; they participate in shaping them within educational discourse. These findings highlight the urgency of fostering critical AI literacy that integrates CLRT with multiliteracies pedagogy to enhance critical and social consciousness where diverse cultural perspectives are not only acknowledged but meaningfully embedded in pedagogy within increasingly AI- mediated learning environments.