Page 376 - CyberSecurityEssentialsEnglish

or review, especially where an internal team would struggle to anticipate the effect of
language, accessibility, or cultural assumptions.

We then describe benefits and harms in measurable terms. A benefit is not a word such as
speed; it is a reduction in average processing time while quality is maintained. A harm is not

a word such as bias; it is a specific scenario: a lower detection rate for a particular group, an
inability to appeal, a data leak, or increased workload on reviewers. Each scenario needs a

plausible cause, an indicator, a control, and an owner.

Record the assumptions and the uncertainty. Results may rest on a small sample, on
historical data that no longer represents the present, on testing in one language, or on
expert users. Rather than hiding these limits, write them as conditions of approval: limited

operation, intensified monitoring, additional data collection, or a prohibition on use outside
a defined scope.

An impact assessment does not end at signature. It is repeated when the purpose, the data,
the model, the level of autonomy, the affected group, or the legal environment changes. It

helps to define re-assessment triggers in advance, such as a major vendor update, a rise in
complaints, a breach of a performance threshold, adding a tool that can act, or moving the
system to another country or sector.

● Remember: an impact assessment is a living decision document

Its value lies in explicit assumptions, specific harm scenarios, testable controls, and dates
for re-review. It is not a public relations report.

Testing in Conditions Close to Reality
A system can succeed in a clean test and fail in operation because of incomplete documents,

dialects, input errors, time pressure, conflicting policies, or questions outside its scope. The
test set must therefore include common cases, edge cases, misleading inputs, contradictory
information, missing data, and situations in which the system should decline to answer or to

act.

Separate the development set from the acceptance set as far as possible. When a team
knows the cases in advance and adjusts the system until it passes them, the test may

improve without generalization improving. Independent reviewers, fresh samples, or blind
testing can be used, with the model version, the settings, and the sources documented so
that a later change in results can be interpreted.

In production, it is not always possible to review every output by hand, but automated

controls can be combined with human samples. Automated controls cover structure, fields,

375