NCA-GENL Question 218
Select 3A company is deploying a Large Language Model (LLM) for customer support but is concerned about ensuring trustworthiness, including bias mitigation, explainability, and data privacy. Which of the following technologies or approaches can be used to enhance the AI's trustworthiness?
- A
Using NVIDIA NeMo Guardrails to enforce ethical and safety constraints in generated responses
- B
Implementing differential privacy techniques to protect sensitive user data during model training
- C
Training the LLM on a larger, uncurated dataset to improve accuracy and diversity of responses
- D
Incorporating SHAP (SHapley Additive exPlanations) to provide insights into model predictions and improve explainability
- E
Relying on prompt engineering alone to control the output of the LLM
Show answer and explanation
Correct answers: A, B, D
Explanation
Enhancing the trustworthiness of AI systems requires a combination of technologies and approaches that address safety, explainability, and privacy concerns. NVIDIA NeMo Guardrails ensures ethical and safe outputs, differential privacy protects sensitive data, and SHAP improves interpretability. These tools and techniques complement each other to build a trustworthy AI system. However, relying solely on larger datasets or prompt engineering is insufficient to address trustworthiness comprehensively.
- A. Correct.
NVIDIA NeMo Guardrails is specifically designed to ensure that LLMs generate safe, ethical, and trustworthy outputs by enforcing safety constraints, which directly contributes to AI trustworthiness.
- B. Correct.
Differential privacy is a widely recognized technique to protect sensitive user information during model training, enhancing trustworthiness by addressing data privacy concerns.
- C. Incorrect.
While training on larger datasets can improve model accuracy, uncurated datasets may introduce biases, which can undermine trustworthiness rather than enhance it.
- D. Correct.
SHAP is a popular tool for explainability in AI models, allowing stakeholders to understand how the model arrived at its decisions, which is a critical aspect of AI trustworthiness.
- E. Incorrect.
Prompt engineering improves control over outputs but is insufficient on its own to comprehensively address issues like bias, explainability, or data privacy.