AIF-C01 Question 76
Select 3A company is using a generative AI model to create customer support content, but the management has noticed that the responses sometimes contain fabricated facts or incorrect information. What are the likely disadvantages of using generative AI in this scenario?
- A
The model may hallucinate and generate content that appears factual but is untrue.
- B
Generative AI models can lack interpretability, making it difficult to understand why specific outputs are produced.
- C
The model is deterministic, meaning the same input always results in the same output.
- D
Generative AI models may not always align with the organization's accuracy requirements.
- E
The model requires significant labeled data for fine-tuning to address these issues.
Show answer and explanation
Correct answers: A, B, D
Explanation
Generative AI models have inherent disadvantages such as hallucination (generating false but plausible outputs), lack of interpretability (difficulty understanding their reasoning), and potential inaccuracy. These issues make them less ideal for tasks requiring high reliability. The nondeterministic nature of these models and the challenges in aligning their outputs with specific requirements further complicate their use in sensitive applications like customer support.
- A. Correct.
Hallucination is a known issue with generative AI models, where they generate outputs that are factually incorrect or fabricated even though they appear plausible.
- B. Correct.
Generative AI models often lack interpretability, meaning it is challenging to understand the reasoning or process behind their outputs.
- C. Incorrect.
This option is incorrect because generative AI models are generally nondeterministic, meaning the same input may produce different outputs due to inherent randomness or variations in the model.
- D. Correct.
Generative AI models may produce outputs that do not meet the required accuracy for specific use cases, which is a major disadvantage in tasks needing reliable data.
- E. Incorrect.
While labeled data can improve model performance, this is not directly related to the disadvantages described in the scenario. The issues of hallucination, interpretability, and accuracy persist regardless of the amount of labeled data.