Advancements in AI Safety and Certification

The field of artificial intelligence is moving towards a greater emphasis on safety and certification, with a focus on developing practical schemes for ensuring that AI systems are safe, lawful, and socially acceptable. This is being driven by the increasing adoption of AI in safety-critical applications, and the need for transparent and reproducible evidence of model quality in real-world settings. Researchers are exploring new approaches to certifying AI systems, including the development of audit catalogs and methodologies for assessing and certifying machine learning systems. There is also a growing recognition of the importance of human oversight in AI systems, and the need to secure this oversight against potential attacks. Additionally, probabilistic model checking is being applied to a wide range of problems, and monitoring of machine learning systems is becoming increasingly important in dynamic production environments. Noteworthy papers include: Secure and Certifiable AI Systems, which presents a framework for assessing and certifying machine learning systems, and Secure Human Oversight of AI, which explores the attack surface of human oversight and outlines hardening strategies to mitigate these risks.

Sources

Safe and Certifiable AI Systems: Concepts, Challenges, and Lessons Learned

An Interval Type-2 Version of Bayes Theorem Derived from Interval Probability Range Estimates Provided by Subject Matter Experts

Secure Human Oversight of AI: Exploring the Attack Surface of Human Oversight

Probabilistic Model Checking: Applications and Trends

Monitoring Machine Learning Systems: A Multivocal Literature Review

Built with on top of