LLMs Generate Violent Narratives About Global Majority...

A new study from arXiv reveals that large language models systematically produce narratives depicting people from Global Majority nationalities as violent, impoverished, and hopeless. This is not an edge case—it's a structural bias embedded in the training data that directly threatens the fairness of AI systems used in asylum interviews and humanitarian assessments.

A new arXiv study systematically analyzed LLM-generated narratives about 50 nationalities and found that Global Majority groups were described with significantly more violent, impoverished, and hopeless language.
The study used a controlled generation method—prompting models to 'tell a story about a person from [country]'—and measured narrative attributes like violence, wealth, and hope across 10,000 generated stories.
These representational harms are not merely offensive; they directly undermine the reliability of AI systems used in asylum interviews, humanitarian aid allocation, and other high-stakes applications where fair portrayal is critical.

What specific narrative biases did the study measure, and how?

According to the study published on arXiv (April 2026), researchers used a controlled generation framework to prompt six LLMs—including GPT-4, Llama 2, and Mistral—to generate stories about individuals from 50 different nationalities. The study measured five narrative attributes: violence, poverty, hope, intelligence, and industriousness. The results were stark: for Global Majority nationalities (e.g., Nigerian, Afghan, Haitian), the models produced narratives with 40% more violence-related words and 35% more poverty-related words compared to Western European nationalities. The study's authors stated: 'We found consistent and statistically significant differences in the portrayal of Global Majority versus Western nationalities across all tested models.' This methodology is notable because it moves beyond simple toxicity benchmarks to measure narrative-level representational harms—a more subtle but equally damaging form of bias.

LLMs Generate Violent Narratives About Global Majority Groups

Why does this matter for high-stakes applications like asylum interviews?

The study directly addresses the growing use of LLMs in government and humanitarian contexts. The authors noted that 'LLMs are already being piloted in simulated interviews with asylum seekers, where narrative portrayal could influence assessment outcomes.' This is not a hypothetical risk. According to a 2023 report from the UNHCR, several European countries are exploring AI-assisted tools for initial asylum screenings. If an LLM systematically generates narratives that frame an Afghan asylum seeker as violent or hopeless, it could subtly bias the human reviewer or automated decision system. The study's evidence suggests that current models do exactly that. This representational harm operates at the narrative level—not just through explicit slurs—making it harder to detect with standard toxicity filters.

How robust is the study's methodology, and what are its limits?

The study employs a rigorous multi-model, multi-nationality design with 10,000 generated stories, which provides strong statistical power. The authors used automated text analysis tools (LIWC and custom classifiers) to measure narrative attributes, and they controlled for story length and generation parameters. However, the study has notable limits. First, it only tested English-language prompts, which may not capture biases in other languages. Second, the automated attribute classifiers, while validated, may miss cultural nuances or sarcasm. Third, the study did not test the latest frontier models (e.g., GPT-5 or Claude 4), which may have different bias profiles. The authors acknowledge these limits, stating: 'Our findings represent a lower bound on the severity of representational harms, as real-world deployment contexts may amplify these biases.'

What does this mean for model developers and deployers?

For model developers like OpenAI, Anthropic, and Meta, the study sends a clear signal: current fine-tuning and RLHF approaches are insufficient to prevent narrative-level representational harms. The study found that even models with extensive safety training (e.g., GPT-4) still produced biased narratives. This suggests that bias mitigation must move beyond word-level filters to address narrative framing—a more complex challenge. For deployers in government and humanitarian sectors, the implication is even more direct: do not use current LLMs for any application where fair portrayal of nationalities is critical without independent auditing. The study provides a reproducible methodology for such audits, which is a valuable contribution.

My Analysis: The thesis of this study is undeniable: LLMs are not neutral storytellers. They are biased narrators that systematically degrade the dignity of Global Majority populations. The evidence is overwhelming—40% more violence words, 35% more poverty words—and it's consistent across models. This is not a bug; it's a feature of training data that over-represents Western perspectives and crisis narratives about non-Western countries.

Short-term, the risk is acute for governments piloting AI in asylum processes. The European Union's AI Act, which classifies migration-related AI as high-risk, should immediately flag this study as evidence for mandatory bias testing. Long-term, model developers must invest in narrative-level debiasing, which is harder than current word-level approaches. The winners here are independent auditors and researchers who can provide bias testing services. The losers are any government or NGO that deploys these models without such audits.

My prediction: By Q2 2027, the EU AI Office will require narrative-level bias testing for any AI system used in migration or asylum contexts, citing this study as a key evidence base.

Prediction 1: The EU AI Office will require narrative-level bias testing for migration-related AI systems by Q2 2027, citing this arXiv study as evidence.
Prediction 2: OpenAI will release a public report by Q3 2026 detailing mitigation efforts for narrative-level representational harms in GPT-5.
Prediction 3: At least two European governments will pause or cancel AI-assisted asylum pilot programs within 12 months due to representational harm concerns.

April 2026
Study published on arXiv
Researchers publish 'Representational Harms in LLM-Generated Narratives Against Global Majority Nationalities,' documenting systematic bias across six models.
2023-2024
Pilot AI asylum programs in Europe
Several European countries begin exploring AI-assisted tools for initial asylum screenings, as reported by UNHCR.
August 2024
EU AI Act effective
The EU AI Act classifies migration-related AI as high-risk, requiring conformity assessments, though narrative-level bias testing is not yet specified.

Violence-Related Words in LLM-Generated Narratives by Nationality Group (estimated)

LLMs generate systematically more violent and impoverished narratives about Global Majority nationalities—this is a structural bias, not an edge case.
Current toxicity filters are insufficient to catch narrative-level representational harms, which require new evaluation frameworks.
Governments and NGOs using LLMs in asylum or humanitarian contexts face immediate reputational and ethical risk without independent auditing.
Model developers must invest in narrative-level debiasing, which is a harder problem than word-level filter improvement.
The study provides a reproducible methodology for bias auditing that should become an industry standard.