Power, Windows, and Dependencies

Exercise: Map it out

Contested Questions in AI Governance

How should we handle open weights as models get more capable?

When, if ever, should frontier development be slowed - and by whom?

[wip] What even are the prospects for international AI governance?

[wip] How much direct control should different governments have over AGI?

[wip] How should we handle the geopolitical implications of catastrophic dual-use capabilities? 

Tracking Hub

Finding Your Leverage Point

More Career Resources

Next Steps: Fellowships

Certificate

State of Play at the Frontier

Evaluating Claims

The Evals Theory of Change

METR

Epoch

The System Card

Beyond Corporate Promises

Exercise: Briefings

Grappling with the Intelligence Explosion 

Reframing the debate

Exercise: Governance Under Acceleration

Visions of Victory

Face off

Argue it out (with yourself)

Optional but highly recommended

**This course is the spiritual successor to Units 4-8 of our old 12-week AI Governance course - Units 1-3 have been absorbed by the [AGI Strategy Course](https://bluedot.org/courses/agi-strategy). You can still access the old curriculum [here](https://bluedot.org/courses/governance/1/1).**

You’re coming in from the AGI Strategy Course - or somewhere equivalent - and know that AI governance matters. Now you will need to show up credibly in governance conversations. That means being able to consume lots of (technical) information critically, assess what it actually tells you, and translate it for people who need to make decisions - fast, accurately, and tailored to what they care about and what would change their mind in a way that matters.

This unit - and this course - work toward building this skill. For Unit 1 you will read the two main sources of capability information (lab disclosures in the form of model cards and independent evaluations), understand what they can and can't tell you, and turn that into briefs that move the right people in the right ways.

But first, let us establish a shared definition of what we mean by (Frontier) AI Governance. This course believes it is the practice of shaping how advanced AI is built and deployed to achieve good outcomes - it does this primarily through policy, institutions, norms, and relationships. It requires both analytical judgment (what would actually work?) and political judgment (what's achievable, and how do I help make it happen?) and is as much a research as it is a implementation effort. 

The field is characterized by a fair amount of disagreement over goals and methods. This makes it even more important to be able to thoroughly evaluate proposals and build the influence to move the ones one believes in. 

Fair warning: This course is intense, you will fight the readings and the clock. It will also teach you - in a [trial](https://ai.stewart.wtf/posts/bluedot-governance/) by [fire](https://heterotopiaai.substack.com/p/parrhesia-praxis-and-poetry-ai-governances) way - to read smart, to use AI well and get you to produce materials that decision-makers might actually use.  

_For feedback and questions, please email the Frontier AI Governance course lead [Josh Landes](https://www.linkedin.com/in/josh-landes12/) at [joshua@bluedot.org](mailto:joshua@bluedot.org). _

_You can find the course session plans (used in discussions) [here](https://docs.google.com/document/d/1MrvxcuyRYF8lWHh9tr7UFvJPUoYAa95GAYiSm5QM7VM/edit?usp=sharing)._

Changelog: 
- 03/14/2026
    - The course was updated to contain 6 units from 5 units after initial testing. 
- 04/05/2026
    - Unit 4 was fully updated 
- 04/13/2026
    - Unit 5 option on "When, if ever, should frontier development be slowed - and by whom?" was added

Planned work
- Switch out Unit 3's Face off proposals
- Add additional contested questions


Throughout this unit - and, hopefully your future career - you'll encounter many claims about AI capabilities and risks. This template gives you an initial method for processing them. Apply it to the readings below and see if you find it useful. 

- Claim
    - What is being asserted? By whom? In what context?
- Evidence
    - What evaluations, baselines, or comparisons support it? How robust is the methodology?
- External validity
    - What does this not tell us about real-world interaction/ability? What gap exists between the test and the deployment context? Where are we looking to get to?
- Threshold relevance
    - Would this change a policy stance, trigger an oversight action, or cross a safety framework threshold? Does it lock us into a certain path?
- Missing info
    - What would you ask for in a hearing, RFI, or expert consultation? What's absent that you'd need to act?



This short explainer lays out a rough evals taxonomy: model safety evaluations vs. contextual safety evaluations, and what each type can and can't tell you. 

**How to read this:** Read it as someone who may soon need to decide whether an evaluation result should change policy.

**As you read, track:**
- What's the difference between knowing a model can do something and knowing it will do a thing? Why and when does this distinction matter? 
- What would you want to know about any evaluation before you trust its results?
- Pay special attention to the external row in the template. This reading is in part about the gap between what evals measure and what the real world looks like.



Now zoom in. This is one specific independent evaluation methodology, examined in depth. METR measures how long a task an AI agent can complete autonomously - and finds this capability has been doubling roughly every 7 months. 

**How to read this:** Where does METR's approach fit on the map you just learned? What does the time horizon metric actually capture? What does it miss? If someone cited this in a debate, what would you want to push back on - and what would you accept?

Don't worry if the methodology section gets technical. Focus on: (1) the core finding and what it implies, (2) the limitations the authors themselves flag, and (3) what this result does and doesn't tell a policymaker.

**As you read, track:**
- What governance-relevant conclusions can you actually draw from this metric? What conclusions would be overreach?
- If capabilities keep doubling at this rate, at what point does it start to matter for governance? Does it already? 
- Check the latest models and how they’re performing! 



The Epoch Capabilities Index (ECI) gives you the big picture (which is closely correlated). It’s a composite metric that aggregates performance across dozens of benchmarks into a single scale, tracking 149+ frontier models over time.

**How to engage with this: **This isn't a something to read end-to-end. Rather, spend 15 minutes exploring.
1. Start with the Benchmarking Hub. Look up where Claude Opus 4.6, GPT-5.2, and Gemini 3 Pro sit relative to each other across different benchmarks. Notice where rankings shift depending on what you measure.
2. Then read the "AI capabilities progress has sped up" data insight. 

As you read, track:
- If you had 30 seconds to explain the ECI to a policymaker, what would you say? What would you leave out?
- How does knowing the cross-model landscape change how you read a single lab's system card?



A system card is how an AI lab tells the world about its own model. It's the primary document you will encounter when trying to understand a newly released AI model. But it's also produced by a company with interests, audiences, and lots of choices about what to emphasize where. 

**How to read this:** Read it as someone who needs to brief someone important tomorrow! Pay attention to these two lenses: 
1. **The content:** What does Anthropic claim this model can do? What evidence do they provide? What evaluations did they/others run, and what did they find? How does what you learned help you assess the evaluation methodology they describe? 
2. **The comms choices:** What gets prominent placement vs. what's more buried? What risks are discussed in detail vs. mentioned in passing? Where is the language precise and where is it vague? What would a careful reader notice is not discussed? 

**As you read, track:**
- What are the 3-5 things a policymaker must know from this document?
- What's the most important thing this document does not tell you?
- What do you think is missing? 
- Run the template from chunk 2 through at least one specific capability or safety claim.



This final resource (written by a course alumna!) shows you what a governance org does with the kind of information we've covered in Unit 1 so far - translating technical assessments into policy recommendations. 

**How to read this:** Read this as an example of the skill you're now also building. Notice how the author takes technical information (capability evaluations, safety frameworks, model security) and turns it into specific, actionable recommendations. What do they assume the reader already knows? Where do they simplify, and does the simplification hold up given what you've read? What information would strengthen - or complicate - their arguments?

**As you read, track:**
- How does this piece use (or fail to use) the kinds of evaluation evidence you've been reading?
- What does the government need to know that isn't in the system card or METR/ECI reports? Where are the information gaps?
- If you were writing this piece, what would you add or change based on what you've read today?



You've now read how a lab presents its model, how an independent evaluator assesses capabilities, and what a governance organization may recommend. Now it's your turn! 

Using what you've learned from the readings, write **three separate quick 150-word briefings** about Claude Mythos Preview's capabilities and risks. Each briefing is for a different person who has plenty of decisions to weigh as a result of this. 


**For a congressional staffer** drafting questions for an upcoming hearing on AI safety.
- They are thinking: What should my boss ask about? What risks should Congress be aware of? What sounds important to constituents? 
- They have 5 minutes to read your brief before a meeting.



Brief 1

**For a policy lead at a competing AI lab** reviewing Anthropic's latest release. 
- What they might want to know: What is Anthropic actually claiming vs. demonstrating? How does this compare to our approach? What does this mean for our safety framework? Are there evaluation choices that are generous to their model?



Brief 2

**For a journalist** writing a piece on the state of frontier AI capabilities. 
- What they care about: What's genuinely new here? What's the story for a general audience? What should readers worry about - or not worry about? Where is the hype and where is the substance? 

