Page 178 - CyberSecurityEssentialsEnglish

Video: Check Yourself with Lateral Reading

How experienced fact-checkers judge a source by
opening other sources rather than by studying the

page in front of them.
Duration: about 10 minutes | Level: Beginner
Learning objective: Apply the four moves to a claim

produced by an AI system.

Source: CrashCourse, MediaWise, and the Stanford
History Education Group [23].

https://www.youtube.com/watch?v=GoQG6Tin-1E
Verified: June 6, 2026

4.8 Bias and Fairness: Not Every Difference Is an Error, and Not Every Consistency Is Justice
◆ Key idea: bias is a social and technical problem.

Harmful bias may arise from the data, from the way the problem is defined, from
measurement, from language, from interface design, from the way the output is used, or
from the surrounding institutional system. Searching for an offensive word is therefore
not enough; you must examine who benefits, who is harmed, and in what context.

Bias in the statistical sense may mean a systematic deviation from the true value, and in the
social sense a pattern that causes unfair treatment or representation. It is important to state
which meaning is intended rather than using the word as a general accusation. Standards

work in this area takes a sociotechnical view: bias does not come from the data alone, but
also from people, institutions, and context [10].

Representation asks: are there enough groups, languages, and cases in the data and in the

testing? A tool may work well with standard Arabic and weaken with a dialect, with certain
names, or with scanned documents, or succeed with a common job title and fail with a local
one. A general average can hide this variation.

Measurement decides what we count as success. If we evaluate a summarization tool by

fluency alone, we may fail to measure the deletion of exceptions. If we evaluate a
recommendation system by clicks alone, we may reward provocative rather than useful
content. The criterion itself can create a bias toward whatever is easy to measure.

Disparate impact occurs when a rule appears uniform but its effects differ between groups.

Testing fairness is not limited to comparing sentences; you must measure rates of error,
rejection, and omission, or the quality of service, across the relevant categories, while
respecting privacy and not collecting data you do not need.

177