Page 179 - CyberSecurityEssentialsEnglish

In everyday tasks, test the output across varied cases: Arabic and foreign names, different
levels of writing, dialects where relevant, ordinary and borderline cases, and requests that
do not match the stereotype. Ask the system about the assumptions on which it built the

result, but do not treat its explanation as proof of its internal mechanism.

+ For the advanced reader: fairness measures can conflict
We may want demographic parity, so that acceptance rates converge between groups; or
equal opportunity, so that detection rates among qualified cases converge; or equalized
odds, so that false positive and false negative rates converge; or calibration, so that a
probability score means the same thing for every group. No single measure fits every

context, and several of them cannot be satisfied at once. The choice depends on the type
of decision, the law, the ethical stakes, and the quality of the data.

TABLE 4-13: SOURCES OF BIAS AND TESTING QUESTIONS

Possible source The question A practical test

Who is absent or under- Compare performance across
The data
represented? representative cases and groups.

Ask for the definition of the
Problem definition and Did we turn a complex concept
variable, the objective, and the
measurement into an incomplete indicator?
alternatives.

The model and the Do language or formatting Test different names, dialects, and
interface change the quality? phrasings.

Has an advisory suggestion Review who takes the decision
The workflow of use
become an automatic decision? and how the output is used.

Does the average conceal a Examine results by category and
The evaluation
recurring harm? by edge case.

✕ Common misconception Correction

Bias may appear in who is excluded, in the

If the system mentions no personal quality of service, in the assumptions
attributes and uses no offensive phrases, made, or in a differing error rate, even
there is no bias. with polite and apparently neutral

language.

178