When an AI impact assessment reveals biased training data, what is the MOST effective method to ensure regulatory compliance?

Retaining the original biased dataset because accuracy is high ignores that high accuracy on biased data reflects and perpetuates discrimination, which regulators explicitly prohibit.

Model transparency helps explain how decisions are made but does not eliminate the biased patterns encoded in the model from unrepresentative training data.

Using diverse and representative datasets to monitor and correct bias is the most effective compliance method because regulatory frameworks such as the EU AI Act require that AI systems produce fair, non-discriminatory outcomes, which cannot be achieved if the model was trained on skewed data. Curating balanced datasets addresses the source of biased outputs rather than masking them through post-hoc adjustments. Ongoing monitoring with updated datasets also produces documented evidence of continued fairness compliance.

Ethical guidelines provide a normative framework but are not a technical remediation of biased data and do not satisfy concrete regulatory compliance obligations.