Appendix 1: Summary of challenges identified in each technical (Section 4) category

Legend Not applicable Dependent Manageable Difficult
Challenges Hand-
crafted
GDO Supervised Unsupervised Generative
Build: Challenges addressed during the design and training (if data-driven) of the system
Output-outcome misalignment (proxy mismatch)
Biased, inequitable or homogeneous outputs Designer responsible Designer responsible
Opacity makes it difficult to understand or review Technique dependent Technique dependent Technique dependent
Input data accuracy assumed Often not known
Confounding variables biases outputs
Value chain hard to interrogate If pre-trained If pre-trained
Problem definition difficult No 'truth' Highly generalisable
Target(s) accuracy assumed No target Often not known
Output accuracy isn't validated by design
Process: Challenges outside the system's build and use relating to how users and the environment interact with the system.</em>
Users vulnerable to automation bias More active controls needed
User hijacking 'Game'-able 'Game'-able if live 'Game'-able 'Game'-able Often arbitrary
Use: Challenges that may arise during the operation of the system
Concept or data drift
Compute cost If DL If DL If DL
Hallucination
Unauthorised data reuse