Page 299 - CyberSecurityEssentialsEnglish

7.1 AI Across the Whole Research Cycle
◆ Key idea: the function of the tool sets the level of verification.
Using AI is safer and more useful when we define its function at each step: is it helping
with discovery, organization, transformation, analysis, or drafting? The closer a step
comes to producing evidence, a judgment, or a decision, the more a person must check it
and the more precisely you must record what was done.

Academic research is not limited to writing a paper for publication. It begins with a problem
or a question, passes through the design of a method, the collection of data or evidence,

their analysis, their interpretation, and then communication about them. Institutional
research also includes program evaluation, policy analysis, service measurement, studies of
beneficiary satisfaction, comparison of alternatives, and preparing decision briefs. In both

cases AI can support the work, but the standard of quality remains the evidence and the way
it was reached.

It helps to divide uses into categories. Exploratory use helps generate initial questions,

synonyms, possible explanations, or a concept map. Transformative use reformats existing
material, such as turning notes into a table, translating an abstract, or extracting fields from
files. Analytical use includes help in writing code, classifying text, or exploring patterns.
Evidential use occurs when a researcher relies on the output to establish a fact or a finding;

this category needs the strongest verification. A further category, decision support, arises
when the output is used to rank policy options or inform a high-impact choice, and it needs

explicit human approval and a decision record. This logic is also useful when reading
classifications of AI systems by purpose and context [28].

⚠ Research warning: an output is not evidence
An AI output does not become evidence unless it can be tied to a source, to data, or to an
analysis that can be inspected. Evidential use is the highest level of risk, so it requires

direct verification of the source, the method, and the context, not merely confidence in
the model’s phrasing.

This does not mean exploratory use is risk-free. A biased initial list may lead to a narrow

question, a translation may hide a terminological difference, and a system may propose a
variable that is easy to measure instead of one that matters. But the danger changes with
the stage. An early idea can be rejected easily, whereas an incorrect number entering a

results table may change the conclusion. The level of review must therefore match the
impact of the error.

AI is useful in reducing administrative friction: organizing files, naming columns, building
templates, preparing a task list, turning meetings into decisions, and summarizing project

progress. This work can raise a team’s productivity without touching the substance of the

298