Page 79 - CyberSecurityEssentialsEnglish

Prompting for Comparison and Decisions
When you use AI to support a decision, the request must not simply be “choose the best.”
The word best is meaningless without criteria, priorities, and constraints. Begin by defining

the available options, the criteria for comparison, the data source for each criterion, and
whether any criterion is a hard requirement that a high score elsewhere cannot make up for.
The cheapest option, for example, may be unacceptable if it fails to meet a minimum

standard of privacy or quality.

Separate four layers in the output: the facts extracted from the sources, the assumptions
the analysis added, the trade-offs between the options, and the conditional
recommendation. This separation prevents a probabilistic opinion from appearing as an

established fact. You can ask the system to write beside each criterion “sufficient data”,
“insufficient data”, or “assumption requiring approval”, and to withhold a final

recommendation when the missing data could change the ranking.

If weights matter, do not ask the model to invent them. Ask it to help you define them or to
test their effect. For instance: “Show the ranking of the options if cost is weighted at 40
percent, then if it becomes 20 percent, and identify when the leading option changes.” In

simplified terms this is sensitivity analysis. Its value is not to hand the decision to the system,
but to show how fragile a recommendation is when the assumptions move.

In decisions that touch people, rights, money, health, or education, write the prompt so that
it asks for alternatives, effects, and uncertainty rather than a binding judgment. The person

with authority must remain identified, and the data, the sources, and the applicable rules
must be reviewed. A good prompt organizes thinking, but it does not transfer ethical or legal
responsibility to the tool.

▣ Practical example: prompting for a comparison that does not hide the
trade-offs
Compare the three options in the table against cost, delivery time, ease of integration,
privacy, and Arabic support. Use only the attached data. For each criterion, state the value,

the source, and any gap or assumption. Do not choose a final option before presenting the
trade-offs. Give a provisional ranking under two different sets of weights, and identify the
data that could change the ranking. Close with a conditional recommendation and the
name of the human decision owner.

Variation and Stability: Why One Successful Attempt Is Not Enough
The same prompt may produce different outputs from one run to the next, especially in
open generative tasks.
This variability is not always a defect; it can help in brainstorming. It becomes a problem

when you need fixed fields, consistent classification, or text that follows a defined policy.

78