← Back to course
Applied AI / Week 02 / Friday
5/6
Week 02 Β· Why It Gets Things Wrong

Friday

Week 2 review
// Confident answers, made-up sources and our habit of trusting machines
⏱ about 15 min

Friday: Week 2 Review

Maya, Ben and Rosa compare their week. Maya's daughter now asks "where did that come from?" about everything.

Ben found one book he could not confirm, and took it off his list.

Rosa rewrote her policy summary herself. "It took ten minutes," she says. "A refund rule I never made would have cost me far more."

This week in five lines

  • Next-word prediction can give right answers, wrong answers and answers that contradict themselves (NIST).
  • Confabulation is false content stated with confidence. People often call it a hallucination (NIST).
  • It is a natural result of how these tools are designed, and it can happen in any output (NIST).
  • Answers can include made-up reasons and citations that make them look trustworthy (NIST).
  • Automation bias is trusting a machine too much. It can make other risks worse (NIST).
REVIEW
  • Read the question.
  • Tap your answer.
Which word does NIST use for a confident false answer?
An answer cites a study you cannot find. What should you do with the claim?
Seeing AI content as better than other sources, without a good reason, is called what?
An answer says "8 groups" and later "6 groups". What is that?
StatementTrue or false?
Confabulation only happens with hard questions.?
Generated content may not say when it is unsure.?
A summary should be checked against the document it came from.?
A confident tone shows an answer is correct.?
WHY THIS EXERCISEThese ideas are the reason every later week asks you to check.
A LINE-BY-LINE FACT-CHECK
  • ?Read the whole answer once.
  • ?Mark every claim: names, numbers, dates, rules and sources.
  • ?Check each claim against a source you trust.
  • ?Fix or remove any claim you cannot confirm.
  • ?Have a person sign off before it is used.
WHY THIS EXERCISEA fixed routine protects you from automation bias.
People often call confabulation a ____.
Before trusting a citation, ____ it and read it.
In your own words, why can a made-up citation be worse than no citation?
WHY THIS EXERCISENIST warns that invented citations can mislead people into trusting a wrong answer.

Sources: NIST AI 600-1, Generative AI Profile (2024)

Tomorrow is a Try It day: fact-check a printed answer line by line.

← Thursday