4.2.4. AI Risk Classification Frameworks
💡 First Principle: Governance effort is finite, so it must be spent where risk is highest. Risk classification sorts AI use cases into tiers based on their potential for harm, then attaches proportionate controls to each tier. This lets low-risk innovation move quickly while high-risk systems get the scrutiny they need, and it gives auditors a clear, repeatable rationale.
The EU AI Act risk tiers are the most widely referenced classification:
| Tier | Description | Examples | Obligations |
|---|---|---|---|
| Unacceptable risk | Practices considered a clear threat to safety or rights | Social scoring that leads to unjustified detrimental treatment, manipulative techniques that cause significant harm, certain biometric uses | Prohibited |
| High risk | Systems with significant impact on health, safety, or fundamental rights | AI in hiring and worker management, credit scoring, education access, essential services, safety components of regulated products | Risk management, data governance, technical documentation, logging, human oversight, accuracy and robustness, conformity assessment |
| Limited risk | Systems where the main concern is people not knowing AI is involved | Chatbots, AI-generated or manipulated content | Transparency obligations (disclose AI interaction, label synthetic content) |
| Minimal risk | Most other applications | Spam filters, inventory optimization, AI in video games | No specific obligations; voluntary codes encouraged |
The NIST AI Risk Management Framework (AI RMF) is a voluntary framework widely used in the US and beyond. Rather than tiers, it organizes risk management into four functions: Govern (culture, policies, accountability), Map (context and identification of risks), Measure (assessment and tracking of risks), and Manage (prioritization and response). NIST also publishes a Generative AI profile that applies the framework to generative AI risks.
Internal risk tiering applies these ideas to an organization's own portfolio. Criteria usually include the impact on individuals, degree of autonomy, reversibility of decisions, data sensitivity, scale, regulatory exposure, and whether the system is customer-facing.
Classification applies across the AI lifecycle. A system's tier should be reassessed when its purpose, data, users, autonomy, or regulatory environment changes. An internal summarization tool that is later exposed to customers, or a recommendation engine that starts making eligibility decisions, has moved tiers.
⚠️ Exam Trap: Uniform governance for all AI is a distractor. Expect the best answer to classify use cases by risk and apply proportionate controls: the strictest controls for systems that affect people's rights, safety, or livelihoods, and lighter-weight controls for low-risk internal tools. Also remember that classification is revisited when the use case changes.
Reflection Question: Classify these three use cases into tiers and name one control each must have: (a) an AI that drafts internal meeting summaries, (b) a public website chatbot answering product questions, (c) a model that ranks applicants for promotion.
