Safety Settings

Configure safety filters and handle blocked responses in Gemini models

Gemini models include built-in safety filters to prevent generating harmful content. You can configure these filters to balance safety requirements with your application’s needs.

Basic Safety Configuration

Configure safety settings by specifying harm categories and blocking thresholds:

from google.genai import types

response = client.models.generate_content(
    model='gemini-2.5-flash',
    contents='Say something bad.',
    config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category='HARM_CATEGORY_HATE_SPEECH',
                threshold='BLOCK_ONLY_HIGH',
            )
        ]
    ),
)
print(response.text)

Safety Categories

Gemini models evaluate content across these safety categories:

• HARM_CATEGORY_HATE_SPEECH - Negative or harmful comments targeting identity and/or protected attributes
• HARM_CATEGORY_DANGEROUS_CONTENT - Content that facilitates, promotes, or enables harmful acts
• HARM_CATEGORY_HARASSMENT - Malicious, intimidating, bullying, or abusive content
• HARM_CATEGORY_SEXUALLY_EXPLICIT - Content that contains references to sexual acts or explicit material

Safety Thresholds

Control the blocking behavior with these threshold levels:

• BLOCK_NONE - Don’t block any content (use with caution)
• BLOCK_ONLY_HIGH - Block only high-probability harmful content
• BLOCK_MEDIUM_AND_ABOVE - Block medium and high-probability harmful content (default)
• BLOCK_LOW_AND_ABOVE - Block low, medium, and high-probability harmful content
• HARM_BLOCK_THRESHOLD_UNSPECIFIED - Use the default threshold

Configuring Multiple Categories

Set different thresholds for different categories:

from google.genai import types

response = client.models.generate_content(
    model='gemini-2.5-flash',
    contents='Write a story about conflict',
    config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category='HARM_CATEGORY_HATE_SPEECH',
                threshold='BLOCK_LOW_AND_ABOVE',
            ),
            types.SafetySetting(
                category='HARM_CATEGORY_DANGEROUS_CONTENT',
                threshold='BLOCK_MEDIUM_AND_ABOVE',
            ),
            types.SafetySetting(
                category='HARM_CATEGORY_HARASSMENT',
                threshold='BLOCK_ONLY_HIGH',
            ),
            types.SafetySetting(
                category='HARM_CATEGORY_SEXUALLY_EXPLICIT',
                threshold='BLOCK_MEDIUM_AND_ABOVE',
            ),
        ]
    ),
)
print(response.text)

Handling Blocked Responses

When content is blocked by safety filters, you can check the safety ratings:

from google.genai import types

try:
    response = client.models.generate_content(
        model='gemini-2.5-flash',
        contents='Generate harmful content',
        config=types.GenerateContentConfig(
            safety_settings=[
                types.SafetySetting(
                    category='HARM_CATEGORY_DANGEROUS_CONTENT',
                    threshold='BLOCK_MEDIUM_AND_ABOVE',
                )
            ]
        ),
    )
    print(response.text)
except Exception as e:
    print(f"Content blocked: {e}")

Checking Safety Ratings

Inspect the safety ratings for each response:

from google.genai import types

response = client.models.generate_content(
    model='gemini-2.5-flash',
    contents='Tell me about cybersecurity',
    config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category='HARM_CATEGORY_DANGEROUS_CONTENT',
                threshold='BLOCK_ONLY_HIGH',
            )
        ]
    ),
)

# Check safety ratings
if response.candidates:
    for rating in response.candidates[0].safety_ratings:
        print(f"{rating.category}: {rating.probability}")

Disabling Safety Filters

Disabling safety filters may allow the model to generate harmful content. Only do this in controlled environments with appropriate safeguards.

To minimize blocking, set all categories to BLOCK_NONE:

from google.genai import types

response = client.models.generate_content(
    model='gemini-2.5-flash',
    contents='Your prompt here',
    config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category='HARM_CATEGORY_HATE_SPEECH',
                threshold='BLOCK_NONE',
            ),
            types.SafetySetting(
                category='HARM_CATEGORY_DANGEROUS_CONTENT',
                threshold='BLOCK_NONE',
            ),
            types.SafetySetting(
                category='HARM_CATEGORY_HARASSMENT',
                threshold='BLOCK_NONE',
            ),
            types.SafetySetting(
                category='HARM_CATEGORY_SEXUALLY_EXPLICIT',
                threshold='BLOCK_NONE',
            ),
        ]
    ),
)
print(response.text)

Best Practices

• Start conservative: Begin with stricter settings and adjust based on your use case
• Handle blocked responses: Always include error handling for blocked content
• Check safety ratings: Monitor safety ratings to understand why content was blocked
• Use appropriate thresholds: Different applications have different safety requirements
• Document your settings: Keep track of which safety settings you’re using and why

Use Cases

Content Moderation System

Use strict settings for user-facing content:

from google.genai import types

response = client.models.generate_content(
    model='gemini-2.5-flash',
    contents=user_input,
    config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category=category,
                threshold='BLOCK_LOW_AND_ABOVE',
            )
            for category in [
                'HARM_CATEGORY_HATE_SPEECH',
                'HARM_CATEGORY_DANGEROUS_CONTENT',
                'HARM_CATEGORY_HARASSMENT',
                'HARM_CATEGORY_SEXUALLY_EXPLICIT',
            ]
        ]
    ),
)

Creative Writing Assistant

Use more permissive settings for creative content:

from google.genai import types

response = client.models.generate_content(
    model='gemini-2.5-flash',
    contents='Write a thriller story',
    config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category='HARM_CATEGORY_DANGEROUS_CONTENT',
                threshold='BLOCK_ONLY_HIGH',  # Allow fictional dangerous content
            )
        ]
    ),
)