Begin with context
A system may be useful for a draft and unsuitable for a decision with substantial consequences. Describe who will use it, what data it processes and who will review its outputs.
Test with real examples
Create a set of typical and difficult cases. Record errors and test again after changes to the model or data. A confident answer is not a measure of its accuracy.
Manage risk as a process
The NIST AI Risk Management Framework is a voluntary resource for considering risks during AI development and use. It emphasises governance, context, measurement and actions to reduce risk.
This is an editorial guide. A specific system needs its own criteria and checks.
