AI research

SnT’s AI research is organised around four complementary themes, covering the full spectrum from foundational methods to applied systems and societal impact. Each theme is a scientific programme, not just a topic area: it has open questions, methods, obstacles, target environments, and expected outputs.

Four AI research themes

Scientific questions & interdisciplinary approaches

How to ensure that AI agents and AI-based cyber-physical systems, such as online assistants and autonomous vehicles, are trustworthy in real-world conditions?
How can we verify that AI systems comply with regulations such as the EU AI Act?
How do we demonstrate reliable operation in deployment environments and characterise failure scenarios before they become operational risks?
How do we design AI systems that are both responsibly developed and genuinely useful without trading one for the other?

Addressing these questions call for an interdisciplinary approach spanning natural language processing, automated testing, requirements engineering, and runtime verification and monitoring, among other techniques.

Obstacles & target environments

Key obstacles include the rapid evolution of AI models and their black-box nature, the absence of robust evaluation strategies for generative content, the limited fidelity of simulation environments, and the challenge of meeting diverse and pluralistic human expectations. Target environments include regulated sectors, such as financial services, healthcare, and public administration, where opacity poses significant compliance risks, as well as cyber-physical systems, such as autonomous driving and space systems, where a lack of reliability can have life-threatening consequences.

Expected outputs

Evaluation frameworks and benchmarks tailored for generative AI, automated testing tools leveraging simulators and generative AI to produce realistic sensor inputs, compliance verification tools, operationalisation guidelines for EU AI governance building on current practices, validated explainability methods, AI-tailored monitoring tools, and roadmaps for AI alignment.