In a company's risk review, the team is evaluating the impact of delivering generative AI output that contains information contrary to fact (a hallucination) as is. Which is the MOST appropriate risk that could arise for end users?
Choosing the risk that delivering misinformation poses to end users.
Users believe the incorrect information and make wrong decisions or take wrong actions.
Correct. A hallucination is the phenomenon in which generative AI plausibly outputs content not grounded in fact. If this is delivered as is, there is a risk that end users believe the incorrect information and make wrong decisions or take wrong actions, suffering harm. It is mitigated by source checking and human review.
The wait time until a response returns grows longer.
Response wait time is a performance (latency) issue and is separate from delivering misinformation.
The risk of delivering a hallucination as is is wrong decisions based on misinformation, so it is incorrect.
The answer to the same question changes each time, causing confusion.
Variation in answers is a user-experience issue caused by non-determinism, a different axis from whether the content is factual.
It is not the risk of believing misinformation as is, so it is incorrect.
Usage charges exceed expectations.
Cost overrun is a cost-management issue and is unrelated to the accuracy of the output content.
It is not the risk that delivering misinformation poses to end users, so it is incorrect.
Delivering a hallucination (output contrary to fact) as is risks 'end users believing the incorrect information and making wrong decisions or taking wrong actions, suffering harm.' It is mitigated by citing sources, grounding, and human review.