Page 107 - CyberSecurityEssentialsEnglish

Chapter Three

Inside the AI Market

How do you choose the right model, tool, and medium for each task?
A practical guide linking the choice to productivity, sources, privacy, cost, and verifiability

Chapter question: the issue is not knowing the names of more tools. It is being able to explain
why you chose a particular system, how you tested it, and what you will do if the model
changes or fails on an important case.

Story: Six Cases That Never Happened
In the spring of 2023, a court in New York was hearing a civil claim against an airline. Counsel
for the plaintiff submitted a brief that looked professional: case names, citation numbers,

and quotations written in confident legal language. But opposing counsel could not find
some of the precedents, and then neither could the court. When copies of the decisions
were requested, it emerged that six of the judicial references had been produced by a

generative tool and did not exist at all. On 22 June 2023 the court imposed monetary
sanctions on the lawyers and their firm personally, and affirmed that the use of artificial

intelligence does not relieve a professional of the duty to verify [1].

The Mata v. Avianca case reveals an error deeper than hallucination. A system skilled at

generating probable text was used as though it were a documented legal database. The
output was well organized, but the function required was the retrieval of real precedents
and proof of their existence. The problem was not merely a short prompt. The type of system

and the method of verification did not match the task. The literature on the risks of language
models notes that statistical fluency can produce convincing language with no guarantee of
meaning or source, which is precisely why verification belongs to the choice of tool rather

than only to a later step [17].

This is why the chapter does not open with the question: which is the best model? The word

best has no practical meaning without a task, constraints, and a success criterion. A tool that
excels at brainstorming may be weak at extracting a complex table. A large model may lose
to a smaller system connected to a trusted knowledge base. And the fastest tool may

become the most expensive once review and correction time are counted.
The aim of this chapter is to build a method of choosing that can be explained and re-

examined. You will learn how to decompose a task, how to understand what lies behind a
product interface, how to distinguish generation from retrieval, computation, and execution,
and how to test a tool inside your own real context, especially when Arabic, privacy, or
sources are part of what success means.

106