AI dependability needs new fault taxonomy: van Moorsel's paper
New taxonomy proposes AI Output Faults to classify model errors beyond hardware/software failures
Computer science researcher Aad van Moorsel has published a paper proposing a fundamental rethink of dependability frameworks for AI systems. The research, titled 'The Uneasy Marriage of AI and Dependability', argues that traditional fault-tolerance taxonomies fail to account for AI's unique failure modes. Van Moorsel introduces the concept of 'AI Output Faults' as a new error class, arguing that incorrect AI outcomes should be treated as distinct failures even when hardware and software operate perfectly.
The paper draws striking parallels between classical dependability mechanisms (like N-modular redundancy and acceptance testing) and AI accuracy enhancement techniques (such as model ensembles and reject options). By proposing a unified taxonomy that treats AI as a third system layer alongside hardware and software, the work aims to establish a common language for evaluating reliability across all components of modern AI systems. The research was submitted to arXiv in May 2026 (arXiv:2608.14564) and represents a significant step toward standardized reliability frameworks for AI deployments.
- Proposes 'AI Output Faults' as a new error classification for incorrect AI model outputs
- Maps traditional fault-tolerance mechanisms (N-modular redundancy) to AI accuracy techniques (ensembles)
- Argues for treating AI as a third system layer (with hardware/software) in dependability frameworks
Why It Matters
This framework could revolutionize how we assess AI system reliability, enabling more robust deployments across critical applications.