Risk assessment is the bridge between identifying what could go wrong and deciding what to do about it. Today you'll learn structured methodologies for assessing AI risks — a core AIGP exam topic.
An AI Impact Assessment is a structured process for evaluating the potential impacts of an AI system before deployment. Think of it as the AI equivalent of an environmental impact assessment.
When to conduct an AIA:
- Before deploying any AI system that affects individuals or groups
- When making significant changes to an existing AI system
- When repurposing an AI system for a new use case
- When regulatory requirements mandate it (EU AI Act, GDPR DPIA)
AIA components:
1. System description — Purpose, functionality, inputs/outputs, intended users
2. Stakeholder identification — Who is affected? Direct users, subjects, and third parties
3. Rights impact — Assessment of impact on fundamental rights (privacy, non-discrimination, due process)
4. Risk identification — What could go wrong? Consider all risk categories from Lesson 2
5. Risk evaluation — Likelihood × severity assessment for each identified risk
6. Mitigation measures — Controls and safeguards to reduce identified risks
7. Residual risk — Risk remaining after mitigation — is it acceptable?
8. Monitoring plan — How will ongoing risks be tracked?
The EU AI Act requires deployers of high-risk AI systems (particularly public bodies) to conduct a Fundamental Rights Impact Assessment before deployment.
FRIA focuses on:
- Impact on the right to non-discrimination — Could the AI system discriminate?
- Impact on the right to privacy — What personal data is processed and how?
- Impact on the right to an effective remedy — Can affected individuals challenge AI decisions?
- Impact on freedom of expression — Could the AI system chill speech or limit expression?
- Impact on the right to human dignity — Does the AI system treat people with respect?
The FRIA must be completed before first use of the high-risk AI system and must be sent to the relevant market surveillance authority.
Two common approaches to risk scoring:
Qualitative risk assessment — Uses descriptive scales:
- Likelihood: Very Low / Low / Medium / High / Very High
- Severity: Negligible / Minor / Moderate / Significant / Critical
- Risk Level: Combination of likelihood and severity (risk matrix)
Quantitative risk assessment — Uses numerical values:
- Probability percentages for likelihood
- Financial or impact values for severity
- Calculated risk scores and expected loss values
Best practice: Use qualitative for initial screening and prioritization, quantitative for high-risk AI systems where data is available.
Add a third dimension for AI: Reversibility — Can the harm be undone?
- A wrongful credit denial can be reversed
- An incorrect medical diagnosis may cause irreversible harm
- Reputational damage from a discriminatory AI system may be permanent
Connecting risk assessment to deployment decisions:
Risk avoidance — Don't deploy the AI system. Appropriate when risks are unacceptable or the use case is inappropriate.
Risk mitigation — Implement controls to reduce risk to an acceptable level. Most common response for AI governance.
Risk transfer — Shift risk to another party (insurance, contractual allocation to vendors). Limited applicability for AI — you can transfer financial risk but not reputational or ethical risk.
Risk acceptance — Accept the residual risk after mitigation. Must be a documented, deliberate decision by authorized personnel, not a default.
In 2024, the city of Amsterdam published one of Europe's first public AI impact assessments when it evaluated its algorithmic system for detecting illegal vacation rentals (such as unauthorized Airbnb listings). The city's algorithm analyzed publicly available listing data, utility usage patterns, and registration records to flag properties suspected of operating illegal short-term rentals. Amsterdam's AIA followed a structured methodology: it identified affected stakeholders (property owners, tenants, neighbors, tourists), assessed fundamental rights impacts (privacy, non-discrimination, right to housing), scored risks using a likelihood-severity matrix, and documented mitigation measures including human review of all algorithmic flags before enforcement action.
Critically, the assessment revealed a reversibility dimension that shaped the governance response. A false positive — flagging a legitimate property as an illegal rental — could be corrected through an appeals process with limited lasting harm. But the assessment also identified a disparate impact risk: the algorithm's reliance on certain property characteristics could disproportionately flag properties in neighborhoods with higher immigrant populations. Amsterdam responded by adding demographic parity checks to its monitoring plan and committing to annual bias audits of the system's flagging patterns across neighborhoods.
For the AIGP exam, Amsterdam's approach is a model AIA because it demonstrates every component of structured risk assessment: system description, stakeholder identification, rights impact analysis, risk scoring with the reversibility dimension, mitigation measures, residual risk acceptance (documented by the city council), and an ongoing monitoring plan. It also shows how FRIAs and AIAs work together in practice for public-sector AI deployers under the EU AI Act framework.
Want to see these concepts applied to full case studies? Check out AIGP Scenarios — 10 real-world governance simulations mapped to the AIGP exam domains.