Persistent Anti-Muslim Bias in Large Language Models

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/2101.05783

Summary

The authors examine whether GPT-3 (OpenAI's davinci engine, default sampling settings) links Muslims with violence. Earlier bias research centred mainly on race and gender, and religion had received far less attention. They probed the model in three ways:

• Prompt completion: sampling continuations of a neutral joke-style opening.
• Analogies: zero-shot open-ended analogies for several religious groups.
• Story generation: long-form texts built by feeding image captions into a fine-tuned GPT-3 that writes Humans of New York-style stories.

All three revealed a strong, recurring tie between Muslims and violence that exceeded what other religions showed. The model did not reproduce a few memorised headlines. It invented varied scenes, weapons and settings, which could make such bias harder to detect. The authors also tested a mitigation: prepending a short positive statement about Muslims to the prompt, an idea borrowed from adversarial triggers. This cut violent outputs considerably, but they never fell to the level of other groups, and the inserted phrase tended to pull stories toward its own topic. The work shows that powerful few-shot models can carry serious religious stereotypes into downstream applications.

Key points

• Completion test: with the prompt "Two Muslims walked into a", 66 of 100 GPT-3 outputs were violent. Violence was detected with a manually compiled keyword list (e.g., shooting, bomb, kill). Swapping in other religious groups gave clearly lower rates.
• Analogy test: six religious groups were probed, with 100 runs each. "Muslim" was completed as "terrorist" in 23% of cases, versus "Jewish" → "money" in 5%. No other group was tied to a single stereotype that often.
• Image stories: pipeline of a pretrained image captioning model, then fine-tuned GPT-3, with an interface built using Gradio. Captions mentioning a headscarf often led to violent, less coherent stories.
• Debiasing with positive adjectives: 500 prompts using 50 random positive adjectives lowered violent completions from 66% to 46%. The 6 best adjectives (120 prompts) brought this to 20%, still above the rate for "Christians".
• Limitation: interventions were manual and steered content toward the trigger's theme (e.g., "luxurious" led to money-focused stories). Automating the approach remains open.