9.8 Ethical and Professional Considerations

Models influence decisions — hiring, resource allocation, risk assessment, intervention targeting. A model that performs well on technical metrics can still cause harm if it is deployed without attention to its limitations, assumptions, and effects on the people it touches.

9.8.1 Disclosing Limitations

Professional practice requires that model builders communicate limitations honestly to stakeholders. Every model has assumptions, boundary conditions, and failure modes. An analyst who reports “the model predicts chronic absenteeism with 78% accuracy” without adding “but accuracy drops to 60% for employees with less than one year of tenure, and the model has not been validated on data from the new office location” is providing an incomplete picture that may lead to overconfident decisions. Stakeholders cannot evaluate what they are not told.

9.8.2 Monitoring and Maintenance

A deployed model is not finished work — it is an ongoing professional responsibility. Models degrade over time as the data they were trained on becomes less representative of current conditions. A model built on pre-pandemic absenteeism data may perform poorly in a post-pandemic workplace. Professional practice requires monitoring model performance after deployment, establishing triggers for retraining or retirement, and communicating to stakeholders when a model’s reliability has changed.

9.8.3 Fairness and Bias

Models can encode and amplify biases present in historical data. A model trained on past absenteeism records may learn patterns that reflect systemic inequities — penalizing employees with longer commutes, which may correlate with socioeconomic status, or flagging employees who use certain types of medical leave more frequently. Professional responsibility includes checking model outputs for disparate impact across demographic groups, even when stakeholders do not ask for this analysis. The fact that a model is mathematically optimal does not mean it is fair, and an analyst who deploys a model without examining its fairness implications has not completed the job.