AI Model Wiki

    Skip to main content
    22 min
    Lesson 2 of 4

    Risk Assessment and Mitigation

    From the course AI Risk Management

    Conduct systematic risk assessments and implement effective mitigation strategies.

    Risk Assessment Process

    Pre-Deployment Assessment

    Step 1: System Characterization
    - What decisions does the AI make?
    - Who and how many people are affected?
    - What data is used?
    - What are the failure modes?
    
    Step 2: Risk Identification
    - Brainstorm failure scenarios
    - Review similar AI incident cases
    - Consult affected stakeholders
    - Red team exercises
    
    Step 3: Risk Analysis
    - Score each risk (Likelihood × Impact)
    - Map risk relationships
    - Identify root causes
    
    Step 4: Risk Evaluation
    - Compare against risk appetite
    - Prioritize high-risk items
    - Determine acceptable residual risk
    
    Step 5: Risk Treatment
    - Avoid, reduce, transfer, or accept
    - Design specific controls
    - Assign ownership
    

    Bias Risk Assessment

    from sklearn.metrics import confusion_matrix
    import numpy as np
    
    def assess_bias_risk(predictions, actuals, protected_attribute):
        groups = np.unique(protected_attribute)
        metrics = {}
        
        for group in groups:
            mask = protected_attribute == group
            tp = np.sum((predictions[mask] == 1) & (actuals[mask] == 1))
            fp = np.sum((predictions[mask] == 1) & (actuals[mask] == 0))
            fn = np.sum((predictions[mask] == 0) & (actuals[mask] == 1))
            tn = np.sum((predictions[mask] == 0) & (actuals[mask] == 0))
            
            metrics[group] = {
                "positive_rate": (tp + fp) / len(predictions[mask]),
                "true_positive_rate": tp / (tp + fn) if (tp + fn) > 0 else 0,
                "false_positive_rate": fp / (fp + tn) if (fp + tn) > 0 else 0
            }
        
        # Calculate disparate impact
        rates = [m["positive_rate"] for m in metrics.values()]
        disparate_impact = min(rates) / max(rates)
        
        # Flag if < 0.8 (80% rule)
        risk_level = "HIGH" if disparate_impact < 0.8 else "LOW"
        
        return {"disparate_impact": disparate_impact, "risk_level": risk_level}
    

    Mitigation Strategies

    Technical Controls

    • Regular model retraining to address drift
    • Robust input validation to prevent adversarial attacks
    • Ensemble methods to improve reliability
    • Uncertainty quantification to know when AI is uncertain
    • Canary deployments for gradual rollout

    Organizational Controls

    • Human-in-the-loop for high-stakes decisions
    • AI ethics review board approval
    • Regular third-party audits
    • Incident response planning
    • AI-specific insurance coverage

    Key Takeaways

    • •Systematic 5-step assessment covers characterization, identification, analysis, evaluation, and treatment
    • •Disparate impact testing identifies bias risk — below 0.8 ratio triggers high-risk flag
    • •Technical controls include drift monitoring, input validation, and canary deployments
    • •Organizational controls with human oversight manage risks that technology cannot fully address
    0 of 4 lessons marked complete

    Progress is kept in this browser only.