________________________________________________________________________________________________ Main
Interview
How can AI be deployed in ways that uphold the highest standards of safety, while also maintaining public trust through transparency and accountability? Deploying AI in a safe and trusted way is about careful design of the system; utilizing assurance frameworks incorporating trust, explainability, safety and security, and impact to govern and guide the development of the algorithms, training sets and acceptance testing for the AI models.
Key deployment considerations:
• Scalable oversight – ensuring that the human has a chance to monitor, intervene or act if necessary. Humans will always be users, creators and authorizers.
• Understanding where AI is a good fit, and how to blend its development and deployment with more traditional engineering practices( such as conventional V & V methods) and how to integrate with current processes smoothly to ensure ongoing compliance with evolving standards and regulations.
• Understanding what the AI is, how it is being integrated – the trusted aspect should also be evidenced with the appropriate tests and tooling that are AI specific( e. g. tests for data poisoning, unwanted bias etc.)
• Monitoring systems – particularly for adaptive AI or where real-world data will gradually shift away from the training set.
• Identifying mismatches between the level of trust in a system and its trustworthiness( whether it merits that trust) and addressing them.
• Using structured assurance trees that cover sources of risk, hazards, undesirable conditions, and mitigation strategies, using an evidence-based approach to make safety arguments within a clearly specified operational design domain.
• Getting buy-in throughout the organization on AI and ensuring there is a diversity of thought when it comes to risks and concerns. tlimagazine. com 11