Fundamentals of Generative AI →hardMultiple Select

Generative AI Leader Fundamentals of Generative AI Practice Question

A team is designing a generative AI application that must avoid generating harmful or biased content. They are considering various techniques to implement safety measures. Which two approaches are recommended for mitigating harmful outputs in generative AI models? (Choose two.)

⚠ Common exam trap

The trap here is thinking that any parameter change, such as increasing temperature or limiting tokens, can serve as a safety measure, when in fact they do not address harmful content.

Answer choices

Why each option matters

Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.

Correct answer & explanation

✓

Using reinforcement learning from human feedback (RLHF) to align model behavior with human values.

RLHF aligns model behavior with human values by learning from human feedback, while content filters provide a post-generation safety net. Together, they form a robust strategy for reducing harmful outputs. Other options like increasing temperature or limiting output length do not effectively mitigate toxicity and may introduce other issues.

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

✗
Fine-tuning the model on a dataset that contains only positive examples.
Why it's wrong here
Training only on positive examples can lead to a model that lacks robustness against negative inputs and may still generate harmful content when prompted adversarially. It does not teach the model to recognize and avoid harmful patterns. A balanced dataset with both positive and negative examples, combined with other techniques, is more effective.
✓
Using reinforcement learning from human feedback (RLHF) to align model behavior with human values.
Why this is correct
RLHF fine-tunes the model using human preferences to reward safe, helpful, and unbiased responses. This aligns the model's outputs with desired ethical standards and reduces harmful content. It is a recommended approach because it directly incorporates human judgment into the training process, making the model more reliable in sensitive applications.
✗
Reducing the max output tokens to limit the amount of generated text.
Why it's wrong here
Limiting output length may reduce the chance of lengthy harmful content, but it does not address the root cause of harmful generation. A short response can still be toxic. This approach is not a reliable safety measure and may harm usability by truncating legitimate responses. It should not be relied upon for mitigating harmful outputs.
✗
Increasing the model's temperature to encourage more diverse and creative outputs.
Why it's wrong here
Higher temperature increases randomness, which can lead to more unpredictable and potentially harmful outputs. It does not mitigate bias or toxicity; instead, it may exacerbate the problem by generating more extreme or off-topic content. Therefore, it is not a recommended safety measure.
✓
Applying content filters that block or flag outputs containing hate speech, violence, or explicit material.
Why this is correct
Content filters act as a post-processing safeguard, scanning generated text for prohibited categories and preventing their display. They are essential for catching harmful outputs that slip through the model. This approach is recommended because it provides an additional layer of defense without retraining the model, and can be updated as new risks emerge.

Go deeper

Generative AI Leader Fundamentals of Generative AI Practice Question

Why each option matters

Option-by-option breakdown

Related to this question