Governance & Safety
A track of P15 · Responsible AI — Ethics, Safety & Governance.
Risk tiers and obligations, LLM failure modes, safety evaluation and incident thinking - governance as a working skill, not a compliance checkbox.
A model in production makes a decision that harms someone. In the room afterwards, three teams each explain why it was not theirs to catch: the modellers evaluated it on the metric they were given, the platform team served what they were handed, and the product owner was told it had passed review. No one is lying. The accountability was never assigned, and that is a design choice made silently, long before the incident.
Governance sounds like paperwork until the day a model in production hurts someone and everyone asks who was supposed to catch it. This track treats governance and safety as a working skill: the judgement to place a system in the right risk tier, spot how it can fail, test for those failures, and respond when one lands.
It opens with the landscape — the EU AI Act's risk tiers, accountability and human-oversight expectations — presented as a map a practitioner uses, not legal text to memorise. Then LLM-specific safety through the responsibility lens: hallucination, prompt injection and misuse categories, framed by what makes each dangerous and who owns the risk (the mechanics live in the LLM pillars; here it is the safety framing). Safety evaluation and red-teaming come next — as governance practice, with the sober counterweight that a passing eval proves far less than it appears to. The track closes on incident thinking: a real case walked end to end — detect, contain, communicate, remediate, prevent — the responsible-AI playbook for the day something goes wrong.