MEPX
Chapter 3 of 10All chapters
  1. 1Does it understand?
  2. 2Does it learn from me?
  3. 3Confidence and truth
  4. 4Detecting AI content
  5. 5Goals and intentions
  6. 6What it is good and bad at
  7. 7Myths about data and training
  8. 8Jobs and the economy
  9. 9Risks, real and imagined
  10. 10Reading AI news

Chapter 3 of 10

Confidence and truth

Fluency is not accuracy.

Hallucination

Confident invented content is a consequence of generating plausible text, not a bug awaiting a patch. Citations, statistics and quotations are the highest risk.

  • Detectors and self-reported confidence do not solve this.
  • Grounding in supplied documents reduces it substantially.

Stated reasoning

An explanation of steps is generated output. It is useful and checkable, and it is not a transcript of internal computation, which is a common misunderstanding.