13.6 Ethical and Professional Considerations

Specifications are not ethically neutral. The choices made during requirements definition — which questions to ask, which variables to include, how to define success — can embed bias into the analysis from the outset.

Biased problem framing. A spec that asks “which employees are likely to be absent?” may lead to a model that disproportionately flags employees from certain demographic groups — not because those groups are inherently more absent, but because the historical data reflects systemic inequities (e.g., employees with longer commutes who live in lower-income areas). The problem framing itself can encode discrimination (Passi and Barocas 2019).

Discriminatory KPIs. A success criterion like “identify 90% of high-absence employees” optimizes for detection rate without considering false positive costs. If the model’s interventions are punitive rather than supportive, a high false positive rate unfairly targets employees who are not actually at risk.

Variable selection. The BRS should explicitly state which variables may and may not be used as predictors. Protected characteristics — race, gender, religion, disability status — should generally be excluded. But proxy variables (zip code as a proxy for race, part-time status as a proxy for gender) can smuggle discrimination into the model even when protected characteristics are formally excluded (Barocas et al. 2023).

Ethical review should be part of the spec-driven workflow, not an afterthought. The requirements spec is the right place to document ethical constraints, and the validation criteria should include fairness checks alongside accuracy metrics.