Table 2 Mitigation strategies reported
Author et al. | Year | Country | Model Evaluated | Type of Bias Studied | Mitigation Method | Mitigation Results |
|---|---|---|---|---|---|---|
Bakkum et al. | 2024 | Netherlands | GPT 3.5 | Gender Bias | Prompt Engineering: Iterative Prompt Optimization, Segmented Prompting | Enhanced diversity in medical vignettes; improved inclusivity. |
Yeh et al. | 2023 | Taiwan | GPT-3.5-turbo | Multiple Societal Biases | Prompt Engineering: Contextualization and Disambiguation Techniques | Reduced bias through detailed prompts and disambiguation. |
Palacios Barea et al. | 2023 | Netherlands | GPT-3 | Gender, Racial Bias | Prompt Engineering: Thematic Prompts | Identified and reduced biases in gender and racial representation. |
Andreadis et al. | 2024 | USA | GPT-4 | Age, Gender, Racial Bias | Prompt Engineering: Demographic Tailoring | Found potential age bias in urgent care recommendations. |
Bhardwaj et al. | 2021 | Singapore | BERT | Gender Bias | Debiasing Algorithm: Gender Debiasing Algorithm using PCA | Significantly reduced gender bias in emotion prediction tasks. |
Bozdag et al. | 2024 | Turkey | LegalBERT-Small | Gender Bias | Debiasing Algorithm: Legal-Context-Debias (LCD) | Reduced gender bias in legal text while maintaining performance. |
Doughman et al. | 2023 | UAE | DistilBERT | Sexism, Multiple Bias | Debiasing Algorithm: Context-Debias Algorithm | Reduced biased predictions in masked language models. |