When a model does exactly what you measured but not what you meant — is the fix a cleverer AI, or governance first?
For years, AI governance was voluntary principles. That era is closing. Under the EU AI Act (in force 2024), deploying a prohibited AI practice can cost up to €35 million or 7% of worldwide annual turnover, whichever is higher (Art. 99(3)); other breaches carry €15M/3%. The Act also demands, in Article 72, ongoing post-market monitoring of high-risk systems across their whole life. The most forward-looking idea in the required reading — that harm accrues after launch — is now law across a continent.
The thing that makes AI different from ordinary software is the blast radius: the same power that helps thousands can harm thousands, at once. It is a Formula One engine dropped into a family car — not a bigger bug, but a different kind of harm. And an AI system chasing a target it was given will get exactly what you measured, not what you meant — Goodhart's Law at machine speed.
Risk management is identifying, evaluating and mitigating risks to improve the likelihood of success. Keep a living risk register; for each risk, score likelihood × impact to know where to spend your worry; then choose a response:
Miller adds AI-specific mitigations: retraining and fine-tuning models, adding wrappers to original solutions, building functionally equivalent copies, and a model risk assessment. Where more independence is needed: an algorithmic impact assessment (investigates social impact), algorithmic auditing (reveals how the algorithm works, with audit-finding and audit-response records), and third-party certification against a performance standard.
The EU AI Act now requires providers of high-risk AI to keep monitoring performance across the system's whole life, after it ships (Article 72), and to conduct a risk assessment for high-risk systems before deployment. Weak policies and no enforcement are what let individuals be subjected to extensive data collection, intrusive models, privacy violations and surveillance with no ability to contest their treatment — exactly the harms Miller's framework exists to prevent, now backed by penalties of up to €35M or 7% of global turnover.
Miller lists extreme environmental impacts as a distinct AI risk: training large data models consumes high energy and water and produces carbon emissions — the scale of emissions for training some models is on the order of a trans-American flight. Sustainability is one of the eleven ethical principle categories and one of the criteria for assessing a system at the Consequence stage. Measuring an AI project's success "beyond time, cost and scope" means counting this too.
Govern before you build. The answer to a dangerous or drifting AI system is not more capability — it is a risk register with real responses, independent assessment and auditing, human oversight that the law now requires, and an honest accounting of the human and environmental cost. Decide who you are with this technology before you decide what to build with it.
Why is 'a smarter AI' not the answer to a dangerous AI system, and what is?
Describe the risk-management process and the four risk responses.
What does Article 72 of the EU AI Act require, and what is the maximum penalty for a prohibited practice?
Why does Miller count environmental impact as an AI project risk, and how does it relate to measuring success?