Compliance

AI Incident Response

Procedures for identifying, containing, and remediating AI system failures or harms.

Definition

AI Incident Response is the systematic approach to managing AI system failures, errors, or harms. Unlike traditional IT incidents, AI incidents may involve:

  • Gradual degradation: Performance decline without obvious failure point
  • Bias emergence: Discriminatory patterns developing over time
  • Hallucinations: AI generating false but plausible information
  • Adversarial attacks: Malicious manipulation of AI systems
  • Cascading failures: AI errors propagating through dependent systems

Key elements of AI incident response:

  • Detection and alerting mechanisms
  • Classification and severity assessment
  • Containment (potentially including AI shutdown)
  • Root cause analysis
  • Remediation and model updates
  • Regulatory notification (where required)
  • Post-incident review

Sources

  • NIST Incident Response
  • EU AI Act Article 73