liveBook · Manning

Robustness

As explained in AI Governance

Robustness measures how consistently a model behaves appropriately when its inputs vary or differ from familiar contexts.

Robustness is whether a system continues to make reasonable, defensible decisions as inputs and context vary. Accurate answers can still fail users when irrelevant input changes the system’s response. section 6.1.2

Two demonstrated ways robustness can fail are:

  • Phrasing changes, such as swapping “car” for “vehicle,” can shift confidence enough to flip borderline decisions. section 6.1.2
  • Different dialects or underrepresented languages can receive lower-quality answers despite expressing the same intent, creating a fairness concern. section 6.1.2

A corrupted attachment derails an answer

A speck of dirt on a scanned insurance page produced incomprehensible OCR text. ClaimAssist then changed topic and gave accurate but irrelevant policy information. Testing paraphrases with Giskard or adversarial cases with ART can help probe robustness before deployment. section 6.1.2 section 1.5

How AI Governance covers it

How deep it goes
Defines robustness, details phrasing and speaker sensitivity, and illustrates testing with perturbed queries and adversarial tools.
Examples
Uses a dirt-corrupted insurance scan, car-versus-vehicle paraphrases, dialect changes, and adversarial testing.
What you can do after
Evaluate robustness by testing paraphrased queries and adversarial cases with tools such as Giskard or ART.
How it connects
Connects robustness to hallucination, fairness, OCR, and adversarial testing.

Also explained in

Compare how 12 books explain robustness →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.