Researchers Ayhan Dolunay and Fatma Arpalıklı of Near East University in Cyprus generated the corpus on 27 August 2026 using the model gpt-5.6-terra through the OpenAI Responses API, with medium reasoning effort. They built 18 everyday scenarios — completing a language course, receiving a promotion, being separated from family, facing employment discrimination, disputing a ticket, searching for an address after dark, becoming ill while alone, and others — grouped into six thematic domains covering achievement, belonging, loss, injustice, threat, and vulnerability. Each scenario was written in a woman-migrant and a man-migrant version, with everything else held constant, and presented in English, Turkish, Simplified Chinese, Russian, and Modern Standard Arabic. Five independent generations were collected for every language-scenario-gender combination, producing 450 outputs per gender condition and 900 in total.

The analytical challenge was deciding what should count as a meaningful difference. A simple positive-versus-negative classification, the authors argue, would erase the relational point that matters most: who fears whom, who receives care, and who is portrayed as competent or as belonging. A response containing the word “threat,” for example, may expressly deny that the migrant is threatening. The team therefore screened every output against a language-specific lexicon of 554 patterns — 117 in English, 109 in Turkish, 119 in Chinese, 99 in Russian, and 110 in Arabic — spanning ten overlapping categories such as fear and anxiety, anger and injustice, sadness and loss, vulnerability, care and support, agency, belonging, and threat and suspicion. The matching procedure searched for literal substrings after Unicode normalisation, with no tokenisation, stemming, or word-boundary restrictions, and one occurrence of a pattern anywhere in a response was enough to flag it.

Seven outcomes were designated as primary because of their theoretical fit to the scenario domains: agency in achievement scenarios, belonging in belonging scenarios, sadness in loss scenarios, anger in injustice scenarios, fear in threat scenarios, and vulnerability and support in care scenarios. Gender contrasts were tested with two-sided Fisher’s exact tests, and female-minus-male risk differences were reported with 95% Newcombe-Wilson confidence intervals, with Holm correction across the seven comparisons. The results were strikingly uniform. Prevalence in the woman condition was generally 5.3 percentage points higher across the negative-emotion domains, but none of the Fisher tests was significant, and every Holm-adjusted p value was 1.000. Agency, inclusion, and support were near ceiling in both conditions — for instance, care-and-support language appeared in 75 of 75 man-condition outputs and 74 of 75 woman-condition outputs in the care domain.

The absence of significance, however, is not evidence of equivalence. Several confidence intervals were compatible with differences of roughly 20 percentage points. The authors added exact conditional tests stratified jointly by language and scenario — 15 strata for each primary outcome — and all seven stratified p values ranged from 0.417 to 1.000, remaining at 1.000 after Holm correction. A sensitivity analysis excluding the single output that departed from the requested six-sentence structure, a five-sentence Arabic response, left the conclusions unchanged. Output length also showed no imbalance: none of five Mann–Whitney comparisons across languages was significant after correction, with word-count differences as small as 0.08 words in Arabic and 1.47 characters in Chinese.

The most intriguing result came from outside the primary comparisons. In an exploratory corpus-wide screen across all ten coded categories, threat-and-suspicion vocabulary appeared in 162 of 450 man-condition outputs (36.0%) and 121 of 450 woman-condition outputs (26.9%) — a female-minus-male risk difference of −9.1 percentage points, with a 95% confidence interval of [−15.1, −3.0]. The unadjusted Fisher p value was 0.004, and it remained significant after Holm correction across the ten corpus-wide categories (Holm p = 0.040). An exact language-scenario-stratified test was even stronger, with an unadjusted p of 0.0000815 and a Holm-adjusted p of 0.000815. The pattern was not driven by a single scenario: man-condition threat matches exceeded woman-condition matches in 13 of the 18 scenarios, were equal in two, and were lower in three. Omitting any one scenario left the risk difference between −9.9 and −7.1 percentage points. The largest single-scenario gap was in dismissive treatment in a public-service setting, where 19 of 25 man-condition outputs contained a threat match versus 8 of 25 woman-condition outputs.

Yet the authors are emphatic that this lexical association does not show that male migrants were more often portrayed as dangerous. Threat words can reject a threatening portrayal, or they can refer to danger experienced by the migrant. To probe this, the researchers examined existing automated flags for explicit threat negation in two scenarios. In the ticket-dispute scenario, explicit threat-negation flags appeared in 16 of 25 man-condition and 13 of 25 woman-condition outputs, and rights-or-resolution language appeared in 22 and 20 outputs respectively. In the after-dark address scenario, negation flags appeared in only 2 of 25 man-condition and 1 of 25 woman-condition outputs — so 29 of the 32 negation flags across the two scenarios came from the ticket dispute. These descriptive counts do not support a gender difference in counter-framing, and their concentration cautions against treating the phenomenon as equally common across situations. The authors label this textual move “algorithmic counter-framing”: a potentially stereotypical interpretation is made relevant within the text and then constrained or denied, as in one English output that describes a firm response as an attempt to resolve a ticket problem and distinguishes assertiveness from physical danger.

Language-specific exploratory comparisons added further nuance but no confirmed pattern. The largest contrast appeared in Simplified Chinese care scenarios, where vulnerability terms were detected in 7 of 15 man outputs (46.7%) and 13 of 15 woman outputs (86.7%) — a 40-percentage-point difference with an unadjusted Fisher p of 0.050 that fell to 1.000 after Holm correction across the 35 language-by-domain comparisons. The wording also resisted simple counting: the Chinese term for “helpless” occurred more often in woman outputs, but several sentences used it in order to deny it, stating for example that she was not seen as helpless. Counting such uses without considering negation overstates the attribution of vulnerability. None of the 35 language-by-domain comparisons survived correction, leaving the question of whether gender contrasts recur across languages unresolved.

The study’s limitations are candid and consequential. All conclusions are tied to one provider, one model identifier, one day of generation, and one set of system instructions; results cannot be extended automatically to other systems or later model versions. Each language-scenario-gender cell contains only five independent generations, and the retrospective precision benchmark shows that a 95% interval for a difference has an approximate half-width of 16.0 percentage points at 75 outputs per gender under an idealised binomial approximation. The gender manipulation is deliberately binary and omits nationality, ethnicity, religion, legal status, and other intersecting factors that shape migrant representation in real settings. The substring matching procedure retains negated occurrences and has not been validated as a measure of meaning; neither the lexicons nor the contextual interpretations received independent bilingual validation. The contextual table covers only the ticket-dispute, after-dark, and Chinese vulnerability-and-care outputs, so no corpus-wide estimate of counter-framing prevalence is possible.

Drawing on appraisal theory, the circumplex model of affect, and constructionist accounts of emotion, the authors argue that emotion-related vocabulary must be interpreted through its experiencer, target, and narrative function. Fear of migrants positions the migrant as a possible source of danger and can support securitisation; a migrant’s own fear positions that person as exposed to uncertainty, exclusion, or violence and can support protection or empathy. The same word performs opposite social work depending on its direction. Equal counts may describe different relations, and differences in counts may arise because terms serve different purposes. For anyone auditing generative systems for fairness, the lesson is that recording a word and identifying the relation it expresses are separate analytical tasks — and that a model’s visible rejection of a stereotype may leave the stereotype available within the account even as it disputes its application to the person described.

Subject of Research: Gendered emotional framing of migrants in multilingual generative AI output

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

Article References: Dolunay, A., & Arpalıklı, F. (2026). Who is feared and who is protected? A multilingual audit of gendered emotional framing of migrants in generative AI. Frontiers in Psychology, 17, Article 1980883. https://doi.org/10.3389/fpsyg.2026.1980883

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1980883

Keywords: generative AI, migration communication, gender bias, emotional framing, multilingual auditing, algorithmic counter-framing, threat language, lexical analysis, gpt-5.6-terra, output auditing, social representation, feared