L02

L02 · in-class activity · 3 cards

Safety, security and trustworthy AI

Trace safety and security concerns, choose a first response and assign operational responsibilities. Compare reasoning across groups and revisit it when new evidence is revealed.

Join your group

Enter the code on the projector.

Lecturer

Opens a room, a projector board and the phase controls.

How it works

Build on L01

You already considered what could go wrong, why, and what control might help. Now use safety and security to examine those concerns more precisely.

Learning goals

L01 already asked what an AI application does, what needs protecting, what could go wrong and why, and what first control could help. L02 makes that analysis more precise: which concerns involve safety, which involve security, and how do they interact? We test whether an explanation survives new evidence. We also connect a first response to the roles that must carry it out. The following lectures develop detailed attacks, defences and governance.

We use the intent-based distinction proposed by Lin, Sun and Shroff in AI Safety vs. AI Security, version 1, especially sections 3–4 and Figure 3 on page 7. It is a useful teaching perspective, not a universally agreed taxonomy or a validated assessment rubric.

First, work independently on one of three self-contained cases. The loading-bay warning is new in L02; forecasting and the document assistant extend patterns introduced in L01. Make five predefined choices: evidence, possible harm, safety/security relationship, first action and responsible owners. Explain how the action could reduce harm, who must act, what remains uncertain and what risk remains if a duty is omitted. The board groups identical choices. Where answers differ, a group explains the difference and a representative of the competing position responds using the case evidence. Revisit those explanations when new evidence is revealed.