Hallucination
Confident but false output from a generative AI model.
Generative models can produce plausible text, citations or figures that are wrong or invented. Controls include grounding answers in approved sources, and human verification before outputs are relied on.
On the exam
When auditors use generative AI, the key control is human verification of outputs against sources before they go into working papers.
Related terms
- Prompt injectionText that tricks a language model into ignoring its instructions.
Test yourself with today’s question of the day.