Reliability and safety

To build trust, it’s critical that AI systems operate reliably, safely, and consistently under normal circumstances and in unexpected conditions. These systems should be able to operate as they were originally designed, respond safely to unanticipated conditions, and resist harmful manipulation. It’s also important to be able to verify that these systems are behaving as intended under actual operating conditions. How they behave and the variety of conditions they can handle reliably and safely largely reflects the range of situations and circumstances that developers anticipate during design and testing.

To ensure reliability and safety in your AI system, you should:

• Develop processes for auditing AI systems to evaluate the quality and suitability of data and models, monitor ongoing performance, and verify that systems are behaving as intended based on established performance measures.
• Provide detailed explanation of system operation including design specifications, information about training data, training failures that occurred and potential inadequacies with training data, and the inferences and significant predictions generated.
• Design for unintended circumstances such as accidental system interactions, the introduction of malicious data, or cyberattacks.
• Involve domain experts in the design and implementation processes, especially when using AI to help make consequential decisions about people.
• Conduct rigorous testing during AI system development and deployment to ensure that systems can respond safely to unanticipated circumstances, don’t have unexpected performance failures, and don’t evolve in unexpected ways. AI systems involved in high-stakes scenarios that affect human safety or large populations should be tested both in lab and real-world scenarios.
• Evaluate when and how an AI system should seek human input for impactful decisions or during critical situations. Consider how an AI system should transfer control to a human in a manner that is meaningful and intelligible. Design AI systems to ensure humans have the necessary level of input on highly impactful decisions.
• Develop a robust feedback mechanism for users to report performance issues so that you can resolve them quickly.

Privacy and security

As AI becomes more prevalent, protecting privacy and securing important personal and business information is becoming more critical and complex. With AI, privacy and data security issues require especially close attention because access to data is essential for AI systems to make accurate and informed predictions and decisions about people.

To ensure privacy and security in your AI system, you should:

• Comply with relevant data protection, privacy, and transparency laws by investing resources in developing compliance technologies and processes or working with a technology leader during the development of AI systems. Develop processes to continually check that the AI systems are satisfying all aspects of these laws.
• Design AI systems to maintain the integrity of personal data so that they can only use personal data during the time it’s required and for the defined purposes that have been shared with customers. Delete inadvertently collected personal data or data that is no longer relevant to the defined purpose.
• Protect AI systems from bad actors by designing AI systems in accordance with secure development and operations foundations, using role-based access, and protecting personal and confidential data that is transferred to third parties. Design AI systems to identify abnormal behaviors and to prevent manipulation and malicious attacks.
• Design AI systems with appropriate controls for customers to make choices about how and why their data is collected and used.
• Ensure your AI system maintains anonymity by taking into account how the system removes personal identification from data.
• Conduct privacy and security reviews for all AI systems.
• Research and implement industry best practices for tracking relevant information about customer data, accessing and using that data, and auditing access and use.

sample page