Who is feared and who is protected? A multilingual audit of gendered emotional framing of migrants in generative AI

Accounts of migration can cast migrants as threatening or vulnerable, recognise their agency, or describe their need for protection and belonging. We examine whether these emotional and social roles differ when a generative model writes about women and men migrants in five languages. We analysed 900 texts produced on 27 August 2026 with gpt-5.6-terra through the OpenAI Responses API, using medium reasoning effort. Each of the 18 scenarios, grouped into six domains, was presented in both gender conditions and all five languages, with five independent generations for each combination. We screened each response against the lexicon for its language. Gender contrasts were examined using Fisher’s exact tests, risk differences with 95% Newcombe-Wilson intervals, and additional exact tests stratified by language and scenario. None of the seven primary domain-congruent comparisons was significant after Holm correction; the intervals remained too wide to establish equivalence. In an exploratory corpus-wide screen, threat-related vocabulary appeared in 36.0% of man-condition and 26.9% of woman-condition outputs (female-minus-male risk difference −9.1 percentage points; 95% CI [−15.1, −3.0]; Fisher p = 0.004; Holm p = 0.040 across 10 categories). The stratified comparison also remained significant after correction. Existing contextual flags identified explicit threat rejection in 29/50 ticket-dispute outputs and 3/50 after-dark address outputs. These focused observations support an exploratory account of algorithmic counter-framing, without establishing its corpus-wide prevalence or a gender difference in its frequency. The contribution is a relational approach to emotional framing: emotion-related vocabulary must be interpreted through its experiencer, target, and narrative function.