An ML project is not justified by model performance or an attractive business case alone. Its purpose, data sources, likely effects, and operation need scrutiny from the perspectives of people affected by it. An otherwise promising result may be inappropriate to use if the data was obtained without clear provenance or consent. section 2.6 section 2.0
Make ethical review a delivery task, not a final check:
- Before committing to the project, review the hypothesis, user stories, and system outline with affected people’s perspectives in mind; seek their direct input where possible. section 2.6
- Assess personal-data impacts and the application’s accountability and governance requirements. A structured tool such as the Canadian government’s Algorithmic Impact Assessment can help expose potential harm, though its questions do not cover every specialist domain. section 2.6
- As the data becomes available, document where and how it was obtained. Revisit ethical, privacy, and security concerns when exploration reveals what the data contains—and what it cannot establish. section 4.5.4 section 6.7
A model that should not determine a person’s future
Predicting whether someone is likely to be a criminal from their appearance illustrates why a technically possible application may be unacceptable. Ask who bears the consequences of mistakes before treating a proposed prediction as a useful business outcome. section 1.2
Record findings early enough to change or stop a line of development. Ethical concerns discovered only after substantial modelling can leave a team with work that cannot responsibly be used, as well as wasted expenditure and potential liability. section 6.0 section 2.0
