Which approach should NOT be used to mitigate hallucination or confidential data leakage?

A Generative Al Engineer has developed an LLM application to answer questions about internal company policies. The Generative AI Engineer must ensure that the application doesn’t hallucinate or leak confidential data.

Which approach should NOT be used to mitigate hallucination or confidential data leakage?
A . Add guardrails to filter outputs from the LLM before it is shown to the user
B . Fine-tune the model on your data, hoping it will learn what is appropriate and not
C . Limit the data available based on the user’s access level
D . Use a strong system prompt to ensure the model aligns with your needs.

View Answer

Answer: B

Explanation:

When addressing concerns of hallucination and data leakage in an LLM application for internal company policies, fine-tuning the model on internal data with the hope it learns data boundaries can be problematic:

Risk of Data Leakage: Fine-tuning on sensitive or confidential data does not guarantee that the model will not inadvertently include or reference this data in its outputs. There’s a risk of overfitting to the specific data details, which might lead to unintended leakage.

Hallucination: Fine-tuning does not necessarily mitigate the model’s tendency to hallucinate; in fact, it might exacerbate it if the training data is not comprehensive or representative of all potential queries.

Better Approaches:

A, C, and D involve setting up operational safeguards and constraints that directly address data leakage and ensure responses are aligned with specific user needs and security levels.

Fine-tuning lacks the targeted control needed for such sensitive applications and can introduce new risks, making it an unsuitable approach in this context.