NCA-GENL Question 200
Select 3An organization is developing a generative AI application that integrates structured tabular data, unstructured text data, and image data. To ensure the application is designed in a manner that is transparent, fair, and verifiable, which of the following practices should they follow?
- A
Implement data preprocessing pipelines to handle biases specific to each data type.
- B
Use explainable AI (XAI) techniques to clarify model predictions for end-users.
- C
Avoid documenting how data is sourced and processed to prevent exposing proprietary information.
- D
Conduct fairness audits across all combined data to identify and mitigate potential biases.
- E
Limit the use of unstructured data to ensure the model remains interpretable.
Show answer and explanation
Correct answers: A, B, D
Explanation
To ensure generative AI applications are transparent, fair, and verifiable, organizations should use robust data preprocessing to handle biases, adopt explainable AI techniques to provide clarity on model predictions, and conduct fairness audits to identify and mitigate potential biases in the data and outputs. Avoiding documentation or limiting data usage without valid reasons compromises transparency and model performance.
- A. Correct.
Data preprocessing is critical to address biases inherent in different data types (e.g., structured, unstructured, and image data). This ensures fairness in the input data and reduces the risk of propagating biases through the model.
- B. Correct.
Explainable AI techniques help improve transparency by providing insights into how the model arrives at predictions, which is essential for building trust with end-users.
- C. Incorrect.
Avoiding documentation of data sourcing and processing undermines transparency, making it difficult to verify the integrity and fairness of the data used.
- D. Correct.
Fairness audits are necessary to evaluate and address biases present in the data and model outputs, ensuring the application is equitable for diverse user groups.
- E. Incorrect.
While interpretability is important, limiting the use of unstructured data unnecessarily restricts the model's capabilities, which is not a best practice.