Table 2.

Illustrations of explainability requirements for different stakeholders and scenarios.

dimension/example	regulation	investigation	service	service	decision support	decision support
stakeholders	regulator	accident investigatora	service provider	end user	expert user	prediction recipient
scenario	system approval	investigate accident or incident	system deployment	service use	decision support	decision support
purpose of explanation	confidence, compliance	clarity, compliance, continuous improvement	confidence, compliance, (continuous improvement)	challenge, consent and control	confidence, consent and control, challenge	challenge
timing of explanations	pre-deployment	post-incident	pre-deployment	same time as decision	same time as decision	same time as decision
data explainability	global	local, global	global	n.a.	local, global	local
model explainability	global (interpretable models, adversarial examples, influential instances)	global (permutation feature importance, counterfactual explanations, TreeSHAP)	global (interpretable models, adversarial examples, influential instances)	n.a.	global (permutation feature importance, interpretable models)	n.a.
prediction explainability	n.a.	local (KernelSHAP, counterfactual explanations)	n.a.	local (KernelSHAP, DeepLIFT, interpretable models)	local (interpretable models, counterfactual explanations)	local (KernelSHAP, DeepLIFT, interpretable models)

^aService Provider may investigate service ‘outages’ (incidents) and Lawyers/Courts may also investigate challenges from decision recipients, using similar methods.