Socio-Ethical Feedback Loops: Integrating Social Science Theories into LLM Fairness Auditing and Correction

Current fairness auditing tends to be technical and metric-driven, often missing deeper social mechanisms (BEATS, Abhishek et al., 2025; Chakraborty et al., 2025). This research direction proposes a novel synthesis: incorporate explicit frameworks from social science—such as stereotype threat, social dominance theory, or intersectionality—into the auditing and mitigation loop. For instance, audits could be guided by scenario-based probes derived from these theories, evaluating how LLMs handle cases of double marginalization or stereotype reinforcement. Negative findings would not only flag bias but also inform the design of targeted counterfactuals, prompt engineering, or adversarial training routines (Babu et al., 2025) that address the root social dynamics. This approach promises both richer diagnosis and more ethically robust interventions, fostering LLMs that better reflect nuanced social realities and ethical commitments.

References:

1. ETHICS AND FAIRNESS IN GENERATIVE AI USING MITIGATING BIAS IN LARGE LANGUAGE MODELS USING ADVERSARIAL TRAINING. Niby Babu, Varghese S. Chooralil, Jucy Vareed, Hrudya K.P. (2025). ICTACT Journal on Soft Computing.
2. BEATS: Bias Evaluation and Assessment Test Suite for Large Language Models. Alok Abhishek, Lisa Erickson, Tushar Bandopadhyay (2025). arXiv.org.
3. “Ethical Ai Through Bias Mitigation In Large Language Models: A Review”. Anindita Chakraborty,, Sampurna Mandal,, Suvojit Mukhopadhyay,, Tiyasa Saha,, Durjay Barman,, Partha Sarothi Roy, Gulshan Kumar Sinha (2025). International Journal of Environmental Science.