Page 43 - CyberSecurityEssentialsEnglish

✕ Common misconception Correction

If the system fails on a simple question, it is Performance varies by task, wording,
useless; if it succeeds on a difficult one, it is context, and tool. Evaluate each use locally
reliable at everything. and set different boundaries according to

the impact of error.

FIGURE 1-6: WHERE EACH LIMIT BITES IN A SINGLE TASK

What to notice: The seven limits are the ones set out in Table 1-7. What this figure adds is
the moment at which each of them appears, because a limit caught late costs the decision,
not a sentence.

Source: Prepared by the author.

● Remember: Section summary

• Key idea: AI offers useful capabilities, not guarantees.
• Practical decision: Match the safeguard to the problem: a source for facts, a calculator
for numbers, and an expert for high-impact decisions.
• Mistake to avoid: Increasing delegation because the output appears reliable or because
the tool succeeded at a different task.

1.8 Why Does AI Succeed at One Task and Fail at Another?

◆ Key idea: Test the tool on the task itself; do not rely on its reputation.
AI may succeed at a complex task and then fail on a simple detail. Researchers describe

this uneven performance as a jagged technological frontier: two tasks that appear similar
to a person can differ greatly in difficulty for a system [9].

42