Ethical risk begins with the nature of the data being used. The sensitivity and origin of training data (option B) determine whether the model risks violating privacy, perpetuating historical bias, or using data that was collected without proper consent.

AAIA identifies data provenance, data sensitivity, and lawfulness of processing as central to ethical AI. Training data that contains sensitive attributes (e.g., health, ethnicity, gender, financial status) must be reviewed carefully for compliance and ethical impacts.

Option A (diverse outputs) relates to performance, not ethics. Option C (update frequency) is part of lifecycle management, not ethical sourcing. Option D (cleaning methods) ensures quality but does not address ethical risk if the underlying data is inappropriate or unlawfully sourced.

Thus, sensitivity and origin of training data are the primary ethical factors.