Fairness in a generative AI system means designing against bias so its outputs do not perpetuate or worsen existing inequalities. Bias can arise from training data, model assumptions, or algorithmic design. Its effects may harm an individual or create distributional harm across a group or society. section 13.1
Fairness has procedural, distributive, and interactional dimensions. It is therefore not only a data-quality problem: technology, society, and policy interact in determining whether a system treats people fairly. section 13.1
Use layered safeguards rather than relying on one fix:
- Train on diverse, representative data, and detect and mitigate bias in both training data and model outputs. section 13.3.3
- Review the model provider’s safety information, test model versions for harms, and retain human oversight. section 13.3.3
- Include responsible-AI checks in output processing. InterpretML, Fairlearn, and IBM AI Fairness 360 are toolkits for explaining model behavior or detecting and reducing bias in ML models. section 10.6
A RAG application
A retrieval-augmented generation system can carry existing training-data biases into its responses. Treat fairness as a deployment concern alongside retrieval and generation errors: assess outputs for bias, maintain oversight, and consider privacy and security when indexed data is confidential. section 7.6 section 13.3.3
Fine-tuning does not remove the obligation to assess fairness. Evaluate a fine-tuned model’s performance and generalization while addressing privacy, fairness, social effects, and applicable guidelines or regulations. section 9.1.2
