Purpose: To ensure that an AI system operates reliably, securely, and predictably under both normal and adverse conditions, and to withstand, detect, and recover from errors, perturbations, or malicious attacks that could compromise safety.
| Organisational / Technical | Measure |
|---|---|
| A. Technical robustness and reliability | |
| TECH | Display confidence scores or uncertainty metrics for each AI-generated safety statement |
| TECH | Use modular and layered design and introduce redundancy and diversity |
| TECH | Map potential component, data, and algorithm failures |
| TECH | Implement a fairness drift detector which triggers model retraining if imbalance occurs |
| TECH | Use sensitivity analysis to check fairness across different input distributions |
| ORG | Conduct ‘disagreement analysis’ sessions – if AI and humans diverge, identify bias sources |
| B. Human-in-the-loop | |
| TECH | Allow users to expand reasoning traces |
| TECH | Provide comparison views between human-validated and AI-suggested reports |
| ORG | Clearly document roles and responsibilities |
| ORG | Schedule formal review workshops, document decisions, include structured disagreement analysis, and involve multiple stakeholders for each safety-critical deliverable |
| ORG | Include diverse expert groups in validating tool outputs |
| C. Traceability | |
| TECH | Ensure datasets are fully traceable (source, preprocessing, labelling, and version control) |
| TECH | Use data audits to detect underrepresented cases (edge scenarios, rare hazards) |
| ORG | Establish SOPs for AI use, version control of datasets and models, validation of AI outputs, audit trails, and periodic reviews |
| BOTH | Implement version control tools to record dataset versions, model weights, code commits, and validation reports |
| D. Safety culture | |
| ORG | Promote a culture of safety-first decision-making |
| ORG | Set up regular training on AI-assisted analysis and decision-making |
| ORG | Conduct regular validation cycles, using benchmark datasets, and expert-labelled cases |
| BOTH | Embed safety reasoning, hazard traceability, and control feedback into every phase of AI development and deployment |
Source: AIOLIA deliverable 3.1