LLMs as a partial social identity actors. Case Study of Pincha La Burbuja® Cyborg Max GPT

This article presents a critical perspective of Cyborg Max GPT, a custom GPT-4 conversational agent developed within the educommunicative project Pincha La Burbuja to analyse affective polarization in #Nakba discourse on X/Twitter. Grounded in Social Identity Theory (SIT), Cognitive Distributed Theories and Critical Discourse Analysis, the study adopts an interpretive qualitative design that conceptualises algorithmic misclassifications not merely as technical failures but as discursive artefacts with sociocultural and pedagogical implications. A corpus of slightly more than 4,000 public posts tagged with #Nakba was collected, of which 2,900 tweets were annotated – five per batch – by Max GPT across six intergroup discourse variables (ingroup/outgroup alignment; emotion toward each group; ingroup bias; outgroup homogeneity bias). A stratified subsample of 450 tweets was manually recoded to identify biases and hallucinations. Using ATLAS.ti co-occurrence queries, four major causal categories—Literal Interpretation, Contextual Bias, Ambiguous Context, and Multimodal Format—were mapped onto group-identity, emotion-related, and bias-related misclassifications. Results reveal systematic vulnerabilities: multimodal and link-dependent inputs produce extensive false negatives in outgroup homogeneity and ingroup emotion; literalism suppresses subtle affective signals; contextual priors inflate bias detection; and ambiguity disproportionately erases latent intergroup meaning. Building on these findings, the article advances a conceptual argument: when applied to discourse in highly polarized environments such as Nakba, LLMs do not operate as neutral observers but as Partial Social Identity Actors. Their patterned misclassifications mirror SIT processes of categorisation, affective labeling, and intergroup bias, suggesting that misinterpretations should not be understood only as computational deficits but as pedagogically valuable artefacts. Errors become formative devices—requiring students to refute, contextualize, and reconstruct meaning using discursive evidence—thus fostering Critical AI Literacy. Moreover, when users delegate interpretation to LLMs, the model acts as a cognitive extension that participates in social categorization and emotional inference, potentially co-constructing distorted representations of conflict—what Osler (s/f) describes as distributed hallucination. For this reason, LLM-based affective polarization analysis—such as Max GPT within Pincha La Burbuja—demands hybrid methodologies, where automated scalability is complemented by qualitative human validation and contextual discourse analysis within Critical Inoculation frameworks. Keywords: Affective Polarization; Social Identity Theory; Distributed Cognition Theory; Critical Media Literacy; Generative AI; Critical Discourse Analysis