When an AI system produces biased outcomes, the instinct is to blame the data. The data was skewed. The training set was unrepresentative. Fix the data, fix the bias.
This framing is convenient and incomplete. Bias enters AI systems through business requirements, feature engineering, labeling decisions, evaluation metrics, and deployment contexts that sit well outside the data science team’s control. The data is often a symptom of organizational choices made upstream.
Seven Sources of Bias
The EIAF identifies seven distinct bias vectors. Historical bias embeds past discrimination into training data. Representation bias occurs when training data does not match the deployment population. Measurement bias arises from proxy variables that encode protected characteristics. Aggregation bias applies a single model to populations with fundamentally different characteristics. Evaluation bias uses metrics that mask disparate performance. Deployment bias uses systems in contexts they were not designed for. Feedback loop bias amplifies initial biases through recursive training.
Each source requires different mitigation. Data augmentation addresses representation bias but does nothing for measurement bias. Fairness constraints in the loss function can address evaluation bias but may worsen calibration. The intervention must match the mechanism.
The Impossibility Theorem
Computer science has proven that certain fairness criteria are mathematically incompatible. You cannot simultaneously achieve demographic parity, calibration, and equalized odds unless base rates are identical across groups. They rarely are.
This is not an excuse for inaction. It is a call for governed decision-making. Someone must choose which fairness criteria apply, document the tradeoffs, and accept accountability. The EIAF requires this choice to be explicit, documented, and approved by the AI Ethics Officer.
The Three-Stage Review
The EIAF mandates bias assessment at three stages: design review before development begins, validation review before deployment, and continuous monitoring in production. Most organizations that discover bias do so reactively, months or years after deployment. The three-stage approach catches it before it reaches production, and monitors for drift after.