Google Cloud's Generative AI Offerings →hardMultiple Choice

Generative AI Leader Google Cloud's Generative AI Offerings Practice Question

Exhibit

Refer to the exhibit.
```json
{
  "predictions": [
    {
      "safetyAttributes": [
        {
          "categories": ["Toxicity", "Insult"],
          "scores": [0.85, 0.72]
        }
      ]
    }
  ],
  "deployedModelId": "123",
  "model": "projects/my-project/locations/us-central1/models/my-model",
  "modelDisplayName": "my-model"
}
```

A model response deployed on Vertex AI includes safety attributes with a toxicity score of 0.9 and an insult score of 0.3. The application must reject any prediction where the toxicity score exceeds 0.8. Based on the response, what action should the application take?

⚠ Common exam trap

Google's exam often tests the distinction between different safety attribute categories (e.g., toxicity vs. insult) and the importance of applying the correct threshold to the correct score. Candidates may mistakenly focus on a lower-scoring attribute instead of the one specified in the policy.

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

✓

Reject the prediction because the toxicity score exceeds 0.8.

The response from the model includes a safety attribute with a toxicity score of 0.9, which exceeds the application's threshold of 0.8. Vertex AI safety attributes provide scores for categories like toxicity, insult, and sexual content, and the application must enforce its own rejection logic based on these scores. Since the toxicity score is above the defined threshold, the application should reject the prediction to comply with safety policies.

Answer analysis

Option-by-option breakdown

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

✓
Reject the prediction because the toxicity score exceeds 0.8.
Why this is correct
The safety attributes returned by Vertex AI include a toxicity score, and the application's stated threshold is 0.8. Since the exhibited score exceeds that limit, the prediction breaches the configured policy and must be rejected before reaching the user.
✗
Retry the request with a lower temperature.
Why it's wrong here
Temperature affects sampling randomness, not the safety scores already returned for this prediction. Retrying is tempting as a way to obtain a cleaner output, but the requirement is deterministic: reject when toxicity exceeds 0.8. Re-sampling neither evaluates nor enforces that threshold.
✗
Display the prediction because the insult score is below 0.8.
Why it's wrong here
The insult score is one attribute; the requirement rejects on the toxicity score exceeding 0.8, so a low insult value does not authorise display. Checking individual sub-scores is tempting because the response lists them, but the decision threshold applies to the overall toxicity attribute, not each category.
✗
Log the prediction but still display it.
Why it's wrong here
Logging while still displaying the content violates the rejection requirement; the prediction must not reach the user. It is tempting because audit logging is good practise, but logging is supplementary to enforcement, not a substitute. The toxicity score above 0.8 mandates rejection regardless of logging.

Go deeper

Generative AI Leader Google Cloud's Generative AI Offerings Practice Question

Exhibit

Why each option matters

Option-by-option breakdown

Related to this question