Week 3 of a four-week AI governance discussion series, alongside the SAIN Utrecht lecture course on AI governance (November–December 2026). Each session builds on the lecture of the week before, but stands on its own, so join one or all.

Week 1 | Why AI governance matters (24 Nov)
Week 2 | Who shapes AI governance (1 Dec)
Week 3 | What policymakers act on (8 Dec)
Week 4 | Europe and the Netherlands (15 Dec)

Evidence or predictions: incidents, system cards and scenarios

The lecture of 2 December followed what policymakers base their decisions on: incidents from the past, such as the July 2026 Hugging Face incident, what labs disclose today in system cards, and scenarios of the future. This week we ask whether the tools that act on that information can actually work, and if not, what other information we need. Guha et al. (2024) assess four common proposals for AI regulation, namely disclosure, registration, licensing and auditing, and find that they share technical and institutional problems, such as unclear standards, limited government expertise and the difficulty of keeping auditors independent.

Questions we'll explore

Should rules wait for incidents, or act on predictions?
What did the system card tell you, and what did you want to know but couldn't find?
Who should check what AI labs disclose?
Of disclosure, registration, licensing and auditing, which would you introduce first, and why?
If these tools can't work yet, what other information would policymakers need, and who could provide it?

Format (90 min): Discussion in small groups and plenary. No presentations, no wrong questions.

Who can join: Anyone curious about AI safety and governance, whether you're new to the topic, a healthy skeptic, or already deep in it. No technical background needed, and no need to have joined the lectures or earlier weeks. You don't need to have answers, just curiosity.

Readings

How to prepare: This group follows the lecture of 2 December. Please reserve about 2 to 2.5 hours, ideally over a few evenings. Did the lecture readings already? Skip Part 1. Short on time? The ★ items (about 1.5 hours) are the essentials.

Part 1 (~55 min): the lecture readings

★ ~10 min: OpenAI, Hugging Face model evaluation security incident
★ ~15 min: GPT-6 Astra System Card, summary and cyber section
~15 min: AI 2040: Plan A, summary
★ ~15 min: Bommasani et al., Advancing Science- and Evidence-Based AI Policy

Part 2 (~35 min): the academic lens

★ ~35 min: Guha et al. (2024), AI Regulation Has Its Own Alignment Problem, George Washington Law Review. Read the introduction (pp. 1475–1480) and Part II on disclosure (pp. 1495–1507).
Short on time? Read the Stanford HAI policy brief on the same paper instead (~10 min)
Optional: Casper, Krueger & Hadfield-Menell, Pitfalls of Evidence-Based AI Policy (~30 min)

Part 3 (~15 min): reflection

★ ~15 min: note one thing the system card disclosed, one thing you wanted to know but couldn't find, and who you think should check it

Attendees consent to filming and photography as participants in this event. By attending, you agree to being filmed or photographed, which may be used for social media, website, and newsletter content. If you wish to attend but do not want to be photographed, please let a member of our staff know during the event so we can accommodate this.

Discussion Group - What policymakers act on

About

Evidence or predictions: incidents, system cards and scenarios

Readings

Also on Tuesday 8 December