NCA-GENL Question 201
Select 3An organization is developing a generative AI model and plans to integrate data from multiple sources, including user-generated content, structured enterprise databases, and real-time sensor data. To ensure the system is designed in a transparent, fair, and verifiable manner, which actions should the organization take during data integration?
- A
Implement a data lineage system to track the origin and transformation of data.
- B
Exclude any data sources that are not in a structured format to avoid potential biases.
- C
Perform bias detection and mitigation techniques on combined datasets before training the model.
- D
Ensure clear documentation of data preprocessing and transformation steps.
- E
Rely exclusively on real-time data for training, as it is less likely to contain biases.
Show answer and explanation
Correct answers: A, C, D
Explanation
To ensure a generative AI model operates in a transparent, fair, and verifiable manner, organizations must adopt best practices during data integration, such as implementing traceability mechanisms (data lineage), detecting and mitigating biases, and documenting preprocessing steps. These actions enhance trust and accountability, while avoiding oversimplified approaches like excluding unstructured data or relying on real-time data alone.
- A. Correct.
Implementing a data lineage system is critical for ensuring the transparency and verifiability of the data being integrated. This helps track the origin and transformations applied to the data, ensuring accountability.
- B. Incorrect.
Excluding unstructured data is not a fair or practical approach, as it may unnecessarily limit the diversity of the dataset and does not directly address issues of bias or fairness.
- C. Correct.
Bias detection and mitigation techniques are essential steps to ensure fairness in the dataset, especially when integrating data from diverse sources that may contain inherent biases.
- D. Correct.
Clear documentation of data preprocessing and transformation steps is necessary to maintain transparency and reproducibility of the data integration process.
- E. Incorrect.
Relying exclusively on real-time data is not advisable, as it may still contain biases and does not address the need for a comprehensive dataset. A balanced approach with diverse data sources is preferred.