Table 1.

Summary of seven factors supporting AI4SG and the corresponding best practices

Factors	Corresponding best practices	Corresponding ethical principle
Falsifiability and incremental deployment	Identify falsifiable requirements and test them in incremental steps from the lab to the “outside world”	Nonmaleficence
Safeguards against the manipulation of predictors	Adopt safeguards which (i) ensure that non-causal indicators do not inappropriately skew interventions, and (ii) limit, when appropriate, knowledge of how inputs affect outputs from AI4SG systems, to prevent manipulation	Nonmaleficence
Receiver-contextualised intervention	Build decision-making systems in consultation with users interacting with and impacted by these systems; with understanding of users’ characteristics, the methods of coordination, the purposes and effects of an intervention; and with respect for users’ right to ignore or modify interventions	Autonomy
Receiver-contextualised explanation and transparent purposes	Choose a Level of Abstraction for AI explanation that fulfils the desired explanatory purpose and is appropriate to the system and the receivers; then deploy arguments that are rationally and suitably persuasive for the receiver to deliver the explanation; and ensure that the goal (the system’s purpose) for which an AI4SG system is developed and deployed is knowable to receivers of its outputs by default	Explicability
Privacy protection and data subject consent	Respect the threshold of consent established for the processing of datasets of personal data	Nonmaleficence; autonomy
Situational fairness	Remove from relevant datasets variables and proxies that are irrelevant to an outcome, except when their inclusion supports inclusivity, safety, or other ethical imperatives	Justice
Human-friendly semanticisation	Do not hinder the ability for people to semanticise (that is, to give meaning to, and make sense of) something	Autonomy