You've spent seven modules learning to recognize where values enter AI systems, how to hold people accountable for decisions they made without realizing they were ethical, and how to demand transparency and structure at every stage.
Now you have to build something that actually works.
This is the capstone. Not a thought experiment or a debate. A real proposal — one you could hand to a city council, a hospital board, a university administration, or a company's leadership team. A governance framework comprehensive enough that if anyone followed it, harm would be substantially harder to hide.
You'll do this by working through a real scenario: a mid-size city about to deploy multiple AI systems with no oversight structure at all. Your job is to propose one before the systems go live. You have a skeptical audience. They want to deploy quickly and efficiently. They'll push back on complexity, cost, and delay. Your framework has to be both rigorous and defensible under that pressure.
Millbrook is a mid-size city in the Mountain West. Population 280,000. Budget-conscious. Tech-forward mayor. The city council just approved $12 million for AI implementation across four critical systems. They're three months from rollout. They have no AI governance framework at all.
The four systems are:
System 1: Predictive Maintenance — AI that analyzes water main sensor data to predict which pipes will fail and schedule repairs before they break, rather than waiting for failures. This is a public safety and infrastructure issue.
System 2: Housing Allocation — An AI system to identify vacant properties at risk of blight, and to prioritize which should enter the city's rehabilitation program. The city has limited resources and wants to maximize impact.
System 3: Permit Processing — AI that pre-screens building permits, flagging high-risk applications for human review and fast-tracking routine ones. Currently, permit processing takes 60 days; the city wants to cut that to 20.
System 4: Budget Forecasting — Predictive models to forecast tax revenue, expenses, and cash flow, allowing the city to make spending and borrowing decisions with more precision.
None of these systems are inherently unethical. All of them could generate genuine public benefit. But all of them will make decisions that affect people's lives — whether a property is eligible for rehab, whether a permit gets delayed, whether the city can afford a program. And none of them have any oversight structure in place before they go live.
The mayor's office has asked you to draft an AI governance framework for the city. You're meeting with the mayor, the city manager, the city attorney, and department heads tomorrow. You need a framework that is rigorous enough to actually catch problems — but also pragmatic enough that they'll actually implement it. If you propose something too complicated or expensive, it won't happen. If you propose something too weak, you're just rubber-stamping the systems as they are.
That's what you're building in the lab: not an idealized governance structure, but a real one. Defensible. Implementable. Specific enough to make a difference.
Effective AI governance has five dimensions. All are essential. None of them should be an afterthought.
Without oversight, AI systems operate in a blind spot. You need a structure that explicitly assigns someone the job of asking uncomfortable questions before deployment. This is usually a committee — a standing body with clear authority to slow down or reject systems that don't meet standards. Oversight must include people with genuine domain expertise (clinicians for healthcare AI, community members for systems that affect vulnerable populations), not just technical and legal staff.
Transparency doesn't mean publishing training data or model weights — some information is genuinely proprietary or involves privacy tradeoffs. But stakeholders and affected communities need to know: what decisions does this system make, what data does it use, how are humans involved in the override process, what happens if it fails. This takes multiple forms: public documentation, community briefings, impact reports, right-to-explanation for affected individuals.
Accountability is the hardest pillar because it requires naming specific people with specific legal and professional exposure if things go wrong. This includes: clear definition of who decided to deploy (with which level of authority), who is liable if the system causes documented harm, what compensation or remedy is available, and who investigates failures.
Systems that affect people must involve the voices of people they affect — not as consultants brought in for a single meeting, but as part of ongoing governance. This means: representation on oversight committees, regular feedback mechanisms, authority to raise concerns without retaliation, translation and accessibility for non-expert participation.
The deployment that's ethical in year one may be unethical in year three as patterns of use shift, as data distributions drift, as social contexts change. Governance must include mechanisms to revisit decisions, retire systems that no longer make sense, and respond to emerging risks. This is not a once-and-done audit. It's a standing obligation.
A governance framework is more than an organizational chart. It's a set of specific, repeatable practices that force the right questions to be asked at the right time. In the lab, you'll build this for Millbrook by addressing four key components.
Who reviews AI systems before deployment? How are they chosen? What authority do they have to slow down, modify, or reject systems? What expertise must they have? What do they report to?
What happens when a city department proposes to deploy an AI system? What information must they provide? What questions does the oversight body ask? What standards must be met? How is disagreement resolved?
What information about deployed systems is made public, and to whom? How do affected individuals and communities learn about systems that affect them? What documentation is required?
What happens if a deployed system causes documented harm? Who investigates? What remedies are available? Who is accountable, and for what?
You'll be building this framework in dialogue with a skeptical city manager who wants deployment to be fast and easy. You'll have to defend every recommendation, explain why it matters, and articulate what risk you're trying to prevent. The framework that emerges won't be perfect — but it will be real, and it will work if anyone actually follows it.