Page 296 - CyberSecurityEssentialsEnglish

Chapter Seven

AI in Research and Scientific Writing

Searching, discovering, synthesizing, and writing without losing scientific integrity
How to use AI to widen research capacity while the question, the evidence, the
interpretation, and the responsibility stay with the researcher
Chapter question: how do we increase a researcher’s capacity to discover and to write

without breaking the chain of evidence?
The central sentence: any claim that does not lead back to a checkable source is a risk,
however scientific the text may look.

The chain of evidence: a question, then a search strategy, then a source, then a claim, then

a synthesis, then disclosure.
Story: A National Policy Undone by Unverified References
In April 2026, South Africa withdrew the draft of its first national artificial intelligence policy

sixteen days after publishing it for comment, once it emerged that reports had found at least
a tenth of its academic references to be fabricated or unverifiable, and apparently generated
by AI. The problem was not a small formatting error. A document meant to propose national

governance institutions, regulatory initiatives, and innovation incentives became open to
doubt because the chain of evidence supporting it could not be inspected [29].

The responsible minister acknowledged that unverified citations were a serious error that
damaged the credibility of the document, and stressed that a person must oversee the work

whenever AI is used in public documents [29]. The department later announced a new path
for reviewing the policy through an independent committee and a revised draft aimed at
public comment [30]. The episode matters because it does not concern a university or a

scientific journal alone; it shows that a policy or institutional text may look organized and
professional and then collapse at the first simple question: do these references actually

exist? And do they say what we attribute to them?

The crisis could have been avoided had a clear verification chain existed: who proposed the
reference? Who opened the publisher’s page? Who matched the identifier or the indexing
data? Who confirmed that the claim in the text goes back to a specific place in the source?
And who recorded the decision? When that chain is absent, AI turns from a research

assistant into a producer of false confidence: measured prose and plausible-looking
references, but with no trail an independent reader can follow.

The story reveals the core of this chapter: AI can accelerate research, but it can accelerate
error too. It can propose a question, keywords, an outline, code, or a summary, but it does

not give any of these the status of evidence automatically. Good research is not measured

295