Question

A law firm uses a large language model to draft internal case summaries. During review, an associate discovers the model produced a summary citing a court case that does not actually exist, though it was worded with complete confidence. The managing partner asks the AI leader to explain what happened and how to reduce the chance of this recurring. What is the best explanation and mitigation?

Accepted Answer

The model hallucinated by generating plausible-sounding but fabricated content; grounding responses in verified source documents and requiring human verification reduces the risk.

Answer

The model ran out of tokens in its context window, so it invented a citation; increasing the maximum output length will eliminate the problem.

Answer

The model's training data was biased against the firm; retraining it on the firm's own cases will guarantee only real citations appear.

Answer

The model experienced a prompt injection attack; adding stricter input filters will prevent any future fabricated citations.