The synthesis of practical measures for each ethics principle allowed us to identify measures that may be applicable across use case contexts, and those that are specific to an AI deployment setting or AI capability (e.g., AI use in a medical vs HR context).
Measures identified across several use cases were technical and organisational in nature.
For instance, oversight mechanisms include the creation of logs; documentation with defined oversight roles to keep humans in the loop; human validation for high-impact decisions or continuous performance validation for traceability; access controls that regulate which data is accessible or which individuals may have access; and knowledge pertaining to the AI system or area of AI application, as well as training to ensure knowledge is available in the organisation.
Use case specific measures addressed requirements or challenges within a domain, such as the need for competitor analyses of AI systems in case of commercial applications (UC5), technical requirements for classified data in AI training or deployment in security contexts (UC4), or specific knowledge of regulations (e.g., regulations for patient data, UC1).
To operationalise AI ethics in the design of AI systems, we provide a summary of relevant technical and organisational measures for each ethics principle identified in the AIOLIA use cases.
- Human oversight
- Autonomy/User agency
- Freedom of expression and non-censorship
- Over-reliance and deskilling
- Robustness and reliability
- Safety/Human safety
- Non-maleficence
- Privacy and data protection
- Transparency and explainability
- Non-bias, fairness and non-discrimination
- Human well-being
- Accountability and responsibility
Source: AIOLIA deliverable 3.1