Accountability assigns and accepts responsibility for an AI system’s operation, effects, security, review, and correction. A disclaimer that AI can make mistakes is not enough: people need to know which data and documents are available, who may access the system, and which constraints apply. section 8.2.1 section 8.6
Make accountability practical by:
- Writing a short policy covering privacy, security, transparency, accountability, fairness, access, and regulatory constraints. section 8.2.1
- Documenting the use case, model development, data sources, preprocessing, output review, and performance-monitoring processes. section 8.2.2
- Assigning responsibility for model changes, infrastructure, communications, data storage, logs, and sensitive information shared with external providers. section 8.6
- Recording verification methods, verification frequency, and any human adjustments or overrides to model outputs. section 8.4.4
Treat the AI like an employee
An AI system needs a job description, defined access to data and resources, performance monitoring, feedback, and intervention when results are misleading or incorrect. Regular review also helps prevent scope creep, such as using an approved model outside its assessed use case. section 8.2.2
Make human decisions visible
If a person changes or overrides an AI output, record that intervention and explain the verification process. This identifies who accepted responsibility for the decision and can reveal problems with the model’s applicability or performance. section 8.4.4
