Artificial Intelligence System Risk Management Methodology Based on Generalized Blueprints

The rapid uptake of artificial intelligence (AI) systems requires similar advances in their governance. Public and private sector institutions want to adopt new AI tools as they perceive potential efficiency gains and value from them. As with every technological advance, the uptake phase of AI is the ideal time to improve the governance, cybersecurity and safety of these systems. The cybersecurity risks in AI systems are similar to the ones in other information technology systems. However, the regulation of AI systems is changing, so new governance tools are needed. Furthermore, the safety and societal impact of AI depends on the technological choices made when building the systems (e.g., biased training data, overfitted machine learning models, model poisoning attacks or needlessly computation-heavy algorithms). AI tools built with large language model technology seem to speak our languages and therefore appear deceptively easy to adopt. The goal of our research is to provide risk management tools that are similarly easy to use, even if they later lead the adopter into setting up a full technical quality management system. We have created three blueprints of AI system deployments to which an organization deploying AI can match their use case. For each blueprint, we have created high-level guidance on which cybersecurity, data rights and ethical aspects the deploying organization needs to consider. Those building AI systems can quickly match their use cases against the blueprints and speed up the secure and ethical adoption of AI.level guidance on which cybersecurity, data rights and ethical aspects the deploying organization needs to consider. Those building AI systems can quickly match their use cases against the blueprints and speed up the secure and ethical adoption of AI.