MEPX
Chapter 5 of 10All chapters

Chapter 5 of 10

What to expect it to get wrong

Predictable failure modes.

Fabrication

Citations, statistics, quotations, dates and names are the highest risk, because plausible-looking versions are exactly what the process produces well.

  • Confidence carries no information about accuracy.
  • Asking the same question twice can yield two different confident answers.

Other weaknesses

Counting, arithmetic it cannot see through, very long documents, and reasoning about the physical world. Tools that let a model calculate or search cover some of this.