Robustness is not about whether a result looks impressive under ideal conditions. It asks whether the central conclusion survives reasonable changes in inputs, assumptions or environment. A model may be highly accurate on a test set yet change its judgement when an image is lightly compressed. Its accuracy may be high while its robustness is weak.
Robustness is not the same as correctness. A mistaken rule can also behave consistently; stability alone does not make a conclusion true. Nor is robustness identical to resilience. Robustness concerns maintaining function through variation, while resilience places more emphasis on recovering after disruption.
The concept therefore shifts attention from a single success to repeatable testing. When assessing an AI output, statistical result or everyday judgement, the most useful question is often not “Was it right this time?” but “Would the conclusion still hold if the wording changed or one reasonable assumption moved?” A judgement that survives such variation is more fit to become part of a lasting knowledge structure.
https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.