- Does the input data contain intrinsic bias?
- Is the bias acceptable? (i.e. if it only contains data about a certain group because this is the scope of the project)
- If not, are there more representative alternatives? Or can we make the dataset more representative?
- Have you thought about which performance metrics are most applicable here to ensure you're not missing bias?
- Does the architecture of the technology particularly susceptible to bias?
More information about this hazard is available via the Data Hazards Project website.