RAIBR02-BP03 Identify potential harmful events impacting robustness
Mishandling foreseeable variations in inputs can create harmful events. Input variations come in two kinds. Intrinsic variations are differences in input data to which an AI system must attend to succeed. Confounding variations are differences in input data that an AI system must ignore to succeed. You should also consider whether slight changes in input data can produce dramatically different outputs and how input instabilities can cascade across system components.
Level of risk exposed if this best practice is not established: High
Implementation considerations
-
Map input variation scenarios and their potential harmful impacts. Consider medical imaging AI, where varying equipment calibrations and scan qualities influence diagnostic accuracy. Document how differences in data format, quality, and characteristics affect reliability.
-
Analyze how input patterns shift over time to identify distribution harms. For example, recommendation systems should adapt to evolving user preferences and emerging content categories. Seasonal trends and special events often introduce unexpected usage patterns.
-
Consider cascading effects in multi-step workflows. In a multi-step AI workflow where one model's output feeds into another, assess how initial inaccuracies could amplify through the chain. For example, in a document processing system, errors in text extraction might affect subsequent classification or summarization steps.
Resources
Related documents:
-
Improve LLM application robustness with Amazon Bedrock Guardrails and Amazon Bedrock Agents
-
ISO/IEC 42001:2023
A.5.4 Assessing AI system impact on individuals or groups of individuals -
ISO/IEC 42001:2023
A.7.4 Quality of data for AI systems
Related tools:
Evaluate the performance of Amazon Bedrock Resources
