What term describes plausible but factually incorrect LLM outputs?

Get ready for the GARP Risk and AI Exam with flashcards and multiple choice questions. Each question comes with hints and explanations. Prepare for success!

Multiple Choice

What term describes plausible but factually incorrect LLM outputs?

Hallucinations describe when a large language model produces text that sounds plausible and fluent but is not grounded in fact or supported by data. The model isn’t intentionally lying; it’s generating what looks like a coherent continuation by predicting next words based on patterns it learned, which can lead to invented statistics, made-up citations, or false details that still feel convincing. This is a common challenge because accuracy and truthfulness aren’t guaranteed by fluency alone, especially for niche topics, numbers, or recent events. This differs from exact match metrics, which judge how closely the output matches a given reference, without considering whether the content is true or grounded. Perplexity is a measure of how likely the model thinks a sequence of words is—more about internal probability and fluency than factual accuracy. Adversarial testing (red-teaming) is a method used to probe model weaknesses and safeguards, not a term for the phenomenon of convincing but incorrect outputs.

Hallucinations describe when a large language model produces text that sounds plausible and fluent but is not grounded in fact or supported by data. The model isn’t intentionally lying; it’s generating what looks like a coherent continuation by predicting next words based on patterns it learned, which can lead to invented statistics, made-up citations, or false details that still feel convincing. This is a common challenge because accuracy and truthfulness aren’t guaranteed by fluency alone, especially for niche topics, numbers, or recent events.

This differs from exact match metrics, which judge how closely the output matches a given reference, without considering whether the content is true or grounded. Perplexity is a measure of how likely the model thinks a sequence of words is—more about internal probability and fluency than factual accuracy. Adversarial testing (red-teaming) is a method used to probe model weaknesses and safeguards, not a term for the phenomenon of convincing but incorrect outputs.

More practice questions