Renaissance LabsLearn

Models · Safety and alignment

Why AI makes things up and how to guard against it

Hallucination, calibration and checking

6 minBeginner9 chapters

The film opens with real cases where invented AI output caused harm. It explains why models hallucinate, including rare facts and benchmarks that reward guessing, introduces calibration and semantic entropy, and looks at retrieval and citations and their limits. It ends with practical habits for checking AI answers.

What to take away

  • A model trained to continue text in a plausible way can produce the right form of an answer with an invented fact inside.
  • The film argues that benchmarks scoring I don't know as zero reward confident guessing, so scoring should penalize confident wrong answers more.
  • Calibration means that when a model is 90% sure, it should be right about nine times out of ten.
  • Retrieval and citations help but do not guarantee truth, so you still need to click the link and read the passage.
  • Be especially careful with names, numbers, dates, quotes and references, and confirm high-stakes answers with a person or original source.

Go deeper

Read the transcript84 lines, click any line to jump

This transcript comes from automatic captions and may contain small errors.

Keep watching