PhD Position AI Alignment: Value Assessment Models & AI Systems

Challenge
:
Developing, operationalizing, quantifying, and embedding complex human and legal values into alignment pipelines for AI systems, open-weights, and foundation models.

Change
:
Advancing from static, generic benchmarks to dynamic, automated validation and red-teaming frameworks tailored for high-risk deployments.

Impact
:
Enhancing police trustworthiness through AI alignment at the Netherlands Police and ensuring compliance with the EU AI Act by engineering measurably aligned AI systems.

AI alignment refers to the goal of making AI systems behave in line with human intentions and values. AI alignment ensures that advanced AI systems operate safely and strictly within the bounds of human intentions, ethical standards, and the prevalent legal frameworks. With the rapid proliferation of AI systems, frontier LLMs, multimodal models, autonomous agents and their growing capability, there are equal risks of misalignment with human, organizational, and societal values through behavioral drift, hallucination, and adversarial exploitation.

Validating models is crucial before decisions can be made about implementation and is important for continuous monitoring of systems in use, and for facilitating effective human oversight of AI. This is particularly important in high-stakes environments like law enforcement. The main challenge is that validation needs to happen simultaneously along a range of different values that are important in a law enforcement context: accuracy, but also fairness, reliability, trustworthiness, and more need to be ensured.

How can we translate abstract democratic, organisational, and societal values such as algorithmic fairness, transparency, explainability (XAI) into rigorous, quantifiable engineering metrics without sacrificing the general utility of said models and AI systems?