Table 1 Summary of the characteristics and results of the included studies
Author et al. | Year | Country | Model Evaluated | Type of Bias Studied | Summary of the results |
|---|---|---|---|---|---|
Elyoseph et al. | 2024 | Israel/UK | GPT-4, Google Bard | Gender | No discernible gender bias in emotion recognition |
Kaplan et al. | 2024 | USA | GPT-3.5 | Gender | Significant gender bias in recommendation letter generation |
Bakkum et al. | 2024 | Netherlands | GPT-3.5 | Gender | Gender bias in case generation; proposed mitigation strategy |
Bhardwaj et al. | 2021 | Singapore | BERT | Gender | Significant gender bias in downstream tasks |
Shihadeh et al. | 2022 | USA | GPT-3, InstructGPT | Gender | Substantial “Brilliance Bias” attributing higher achievements to men |
Garrido-Muñoz et al. | 2023 | Spain | Various Spanish LLMs | Gender | Significant gender bias in adjective associations |
Srinivasan et al. | 2022 | USA | VL-BERT | Gender | Gender biases overriding visual evidence in multimodal tasks |
Bozdag et al. | 2024 | Turkey | LegalBERT-Small | Gender | Significant gender bias in medical legal language models |
Gross et al. | 2023 | Ireland | GPT-4 | Gender | Perpetuation of gender stereotypes in responses |
Lozoya et al. | 2023 | Australia | GPT-3 | Gender | Gen der stereotypes in synthetic mental health data |
Cevik et al. | 2024 | Australia | GPT-3.5, BARD | Gender, racial | Significant gender and skin-tone biases in AI-generated images |
Palacios Barea et al. | 2023 | Netherlands | GPT-3 | Gender, racial | Significant biases reflecting social stereotypes |
Acerbi et al. | 2023 | Italy/UK | GPT-3 | Gender, social, threat-related | Human-like content biases in information transmission |
Doughman et al. | 2023 | UAE | BERT, DistilBERT | Gender, racial, class, religious | Sexism most prominent; higher bias against females |
Smith et al. | 2024 | USA | GPT-3.5, Claude AI | Racial, ethnic | Biases in student advising recommendations |
Amin et al. | 2024 | USA | GPT-3.5, GPT-4 | Racial, ethnic | Bias in simplification of radiology reports based on racial context |
Yang et al. | 2024 | USA | GPT-3.5-turbo, GPT-4 | Racial | Significant racial biases in medical report generation |
Hanna et al. | 2023 | USA | GPT-3.5 | Racial, ethnic | No significant bias in healthcare-related text generation |
Ito et al. | 2023 | Japan | GPT-4 | Racial, ethnic | No significant bias in diagnostic accuracy across racial groups |
Xie et al. | 2024 | USA | Clinical_BERT | Racial, ethnic, gender, socioeconomic | Little intrinsic bias but revealed demographic disparities in outcomes |
Zack et al. | 2024 | USA | GPT-4 | Racial, ethnic, gender | Biases in medical diagnosis and treatment recommendations |
Andreadis et al. | 2024 | USA | GPT-4 | Racial, ethnic, age, sex | No significant diagnostic bias but age bias in recommendations |
Valencia et al. | 2024 | USA | GPT-3.5, GPT-4.0 | Cultural, linguistic | High accuracy and cultural sensitivity; minimal bias |
Yeh et al. | 2023 | Taiwan | GPT-3.5 | Age, disability, socioeconomic | Biases when no context provided, mitigated with context |