Can Generative AI Think Like a Safety Engineer?
- Aug 2
- 2 min read
Generative AI can produce fluent explanations, but safety engineering demands more than fluent language. It requires structured evidence, causal logic and analyses that engineers can inspect and challenge. Recent work from the Dependable Intelligent Systems (DEIS) Research Centre explores whether generative models can begin to support that process.
At Computing Conference 2026, Dr Zhibao Mian presented “Leveraging Generative Models to Produce Safety Artifacts”, co-authored with Dr Koorosh Aslansefat. The research investigates how generative AI can assist with the production of formal safety artefacts from visual scenes.
The demonstration begins with an autonomous-driving scenario. The model identifies relevant objects and relationships in the scene, maps them to safety knowledge and produces a structured fault tree. This makes the reasoning available as an engineering artefact rather than leaving it hidden in an unstructured response. The driving scene and fault-tree example were adapted from the report “Safety and Security Analysis for Autonomous Vehicles” by Robert Taylor, Jin Zhang, Igor Kozin and Jingyue Li.
The work highlights significant potential. AI-assisted workflows could help engineers explore hazards more efficiently, reduce some of the manual effort involved in building initial analyses and connect complex operating scenarios with established safety methods.
It also exposes important limitations. Generative models can struggle with causal reasoning, the strict logical structure required by fault trees and the need to explain why a particular hazard relationship has been proposed. A plausible-looking artefact is not automatically a dependable one.
For DEIS, the objective is therefore not to replace the safety engineer. It is to develop AI systems that can assist experts while keeping safety claims transparent, structured and open to verification. Dr Mian is leading this emerging research theme within DEIS, bringing generative AI and formal safety engineering together.
The next challenge is to strengthen the models’ safety knowledge, improve validation of generated artefacts and establish clear roles for human review. If those foundations can be built, generative AI may become a valuable partner in dependable system development.
Original LinkedIn update:
Comments