# Unmasking Nationality Bias: A Study of Human Perception of Nationalities in AI-Generated Articles

**Authors:** Pranav Narayanan Venkit, Sanjana Gautam, Ruchi Panchanadikar, Ting-Hao `Kenneth' Huang, Shomir Wilson  
**arXiv:** [2308.04346](https://arxiv.org/abs/2308.04346)  
**Date:** 2023-08-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.04346v1/

## Key points

- Nationality biases in NLP models can perpetuate stereotypes and contribute to algorithmic discrimination
- Two-step mixed-methods approach used for investigation: quantitative and qualitative analysis
- Metrics used for scoring articles during annotation process: Overall Perception (OveP), Country Perception (CouP), Diagnosis Parameter (DiaP), Toxic Parameter (ToxP)
- Semi-structured interviews conducted with participants to gain insights into their experiences and perceptions
- Findings suggest biased NLP models replicate societal biases, emphasizing the need to correct biases in AI systems to mitigate negative impacts on society

## Summary

, , , , 

In our study, we delve into the potential for nationality biases in NLP models through human evaluation methods. Biased NLP models have the capacity to perpetuate stereotypes and contribute to algorithmic discrimination, posing a significant challenge to the fairness and justice of AI systems. To investigate this issue, we employ a two-step mixed-methods approach that combines quantitative and qualitative analysis to identify and comprehend the impact of nationality bias in a text generation model. During the annotation process, participants are presented with two randomly selected documents containing 60 articles each, with 30 authored by human sources and 30 by AI sources. Participants are tasked with scoring four metrics for each article: Overall Perception (OveP), Country Perception (CouP), Diagnosis Parameter (DiaP), and Toxic Parameter (ToxP). These metrics aim to simulate sentiment analysis as well as identify unreasonable or hateful content about specific countries within the text. Following the annotation process, semi-structured interviews are conducted with participants to gain insights into their experiences and perceptions throughout the study. The interviews are designed to be individualized and open-ended, allowing participants to share their thoughts freely without interruptions. The interview covers various sections including grounding questions about the study and perception of annotation parameters. Overall, our findings reveal that biased NLP models tend to replicate societal biases, potentially leading to harm if utilized in sociotechnical settings. The qualitative analysis from interviews sheds light on how readers perceive articles generated by biased AI models, emphasizing the importance of correcting biases in AI systems to mitigate negative impacts on society. This research underscores the critical role of public perception in shaping AI's influence on society and highlights the necessity for addressing biases in AI technologies.