Intro course graduates
You have done AI Risk Fundamentals or a similar course and want to understand one of the field's main agendas properly.Go deeper on one of the main agendas.
About this course
Alignment tries to make an AI model want what we want. AI control asks a different question: if we cannot be sure a model is not secretly working against us, can we still use it safely?
In this course you build a basic understanding of AI control: what it is, its key terms, and its theory of change, so you can form your own view on whether to work on it.
You need a basic understanding of AI safety, from AI Risk Fundamentals or a similar course. No background in control needed.
Who is it for:
People choosing what to work on
You are deciding where to put your effort in AI safety and keep hearing about control.Decide whether control is for you.
The curious
You have heard of AI control but do not know much about it yet.Learn what control is and why it matters.
Anyone who wants the map
You want an overview of the field, its people and its debates, whether or not you later go into technical control work.Get an overview of the field.
Curriculum
5 UNITS · ~3.5–4H EACHEach unit is reading, videos and exercises on your own, with Lens Tutor, then a short test and a group meeting.
What control is
The core idea and its vocabulary, how it differs from alignment, trusted and untrusted models, how a control evaluation works, and what AI companies already run. It ends with a first criticism.
The key ideas and the field
High-stakes and low-stakes threats, the main research areas, and the field itself: who works on control, and how big it is next to the rest of AI safety.
Why people work on control
Control's theory of change: the routes by which it is meant to reduce risk, and how each depends on the political will to act on AI risk.
Criticisms of control
The main criticisms, sorted by what they dispute, including the argument that control might make things worse.
Debates, and your own view
Real debates between people who disagree, including those after the 2026 Hugging Face incident. Find the cruxes, then state and defend a view of your own.
Course Outcomes
Control versus alignment
Tell a control argument (it could not cause harm even if it tried) from an alignment argument (it will not try), and say why the first can be tested today.
Reading a control setup
Say who monitors whom, which models count as trusted, what a setup protects against, and where an attacker would try to get through.
High stakes and low stakes
Tell threats where one action can be a catastrophe from those where harm builds up over many actions, and say why the defence differs.
A map of the field
The main organisations and people, the field's rough size, and its main research areas.
The case and the criticisms
Lay out how control is meant to reduce risk and where each step could fail, and sort the criticisms by what they dispute.
A view of your own
Find the crux in a debate about control, and state your own calibrated view and what would change it.
Upcoming cohorts
INTENSIVE
5h/day · 1 week
Dates to be announcedWe’ll publish the next intensive cohort here.
PART-TIME
5h/week · 5 weeks
Dates to be announcedWe’ll publish the next part-time cohort here.
Common questions
Not sure it’s for you? Come to the first session. If it’s not the right fit, we’ll help you find something that is.