引導式測驗 · 中等 等級

為什麼 AI 聽起來很有自信,即使它是錯的

Explain confident confabulation, why fluent wording is not proof, and how to check citations, uncertainty and consequential answers.

問題 1 的 8

A chatbot gives a polished explanation and a citation, but the cited book cannot be found in a library catalog. What is the main lesson?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. A chatbot gives a polished explanation and a citation, but the cited book cannot be found in a library catalog. What is the main lesson?
    • Fluent reasoning and citation formatting do not prove a source exists
    • A specific title is reliable because it is detailed
    • The library catalog is unnecessary for verification
    • The chatbot must have read the book previously
  2. How does NIST describe confabulation in generative AI?
    • A guarantee that a system checks every fact
    • A special tool for web searching
    • Erroneous or false content that can be presented confidently
    • A type of encryption applied to model output
  3. A chatbot states an exact return deadline but does not link to current policy. What should a support agent do?
    • Accept the date because it is precise
    • Open the current official policy and verify the deadline
    • Ask the chatbot to say it more confidently
    • Average the date with an older answer
  4. Why can a model generate a convincing answer when information is missing?
    • It has verified every statement against a live database
    • It can only produce facts that appear in the prompt
    • It refuses to answer without perfect information
    • It predicts plausible text from learned patterns and the prompt context
  5. A user asks for a confidence percentage and receives “98% sure.” What does that number establish by itself?
    • Only that the model produced that number; calibration must be evaluated separately
    • That 98 out of 100 similar answers are correct
    • That a source was independently checked
    • That the model used a deterministic calculation
  6. Which check is best for an AI-generated calculation?
    • Inspect whether the answer uses confident wording
    • Ask for a longer explanation only
    • Recompute it independently or use a suitable calculation tool
    • Assume arithmetic is correct if formatted neatly
  7. What can a request to list sources improve, and what can it not guarantee?
    • It guarantees that every reference is genuine
    • It can make evidence easier to inspect, but each citation still needs verification
    • It makes external checking unnecessary
    • It proves the answer is current
  8. When should a fluent AI answer not be used as the final authority?
    • When a user asks for a creative story
    • When the answer is short
    • When the model uses plain language
    • When a consequential decision needs authoritative evidence or accountable expertise