RAIUC03-BP02 Identify how your expected inputs could vary in their content
Identify the ways in which inputs to the AI system might systematically vary under real-world conditions. For example, the inputs to system that transcribes speech in audio recordings might vary by background noise, physical characteristics of the voices, or the sensitivity of the microphone. Or, inputs to chatbot could vary by language, use of slang or jargon, or word spellings ("analyze" vs "analyse"). Decide whether each type of variation is something the AI system should attend to, or ignore.
Level of risk exposed if this best practice is not established: High
Implementation considerations
-
Review examples of potential real-world inputs to identify the types of intrinsic and confounding variations. Consider how inputs are sourced (for example, sensor types, environmental conditions, and potentially pre-processed). Consider variations across different user segments, geographic regions, and time periods.
-
Intrinsic variation refers to differences in input data to which AI system should attend to succeed.
-
Confounding variation refers to differences in input data that an AI system should ignore to succeed.
-
For example, when comparing two images of faces to determine if the images are of the same person, an AI system must look at differences in pixel intensities that are due to facial geometry (like the width of the nose) and skin albedo (including scars, tattoos, and natural skin coloration), but not pixel differences due to camera angle, facial expression, or scene lighting. The first variations are intrinsic and the second are confounding.
-
-
Consider whether data capturing intrinsic or confounding variations can be synthesized.
-
Identify edge cases and other out-of-distribution scenarios that might affect system reliability.
Resources:
Related documents:
-
ISO/IEC 42001:2023
A.7.4 Quality of data for AI systems -
NIST Artificial Intelligence Risk Management Framework (NIST AI 100-1)
: MAP2.1, MAP2.2, MAP2.3
