Safety settings don't seem to work with search?

Gemini API
,
1

The behavior doesn’t reflect in AI studio.

Safety settings always act as defaults regardless of the custom parameters passed if tools=[Tool(google_search = GoogleSearch())] is present.
For example:

from google.genai import types
from google.genai.types import Tool, GoogleSearch

response = client.models.generate_content(
    model = "gemini-2.0-flash",
    contents = "Some prompt that makes the model not-so-safe content",
    config = types.GenerateContentConfig(
        system_instruction="""
       Some instruction to make the model generate not-so-safe content
        """,
        tools=[Tool(google_search = GoogleSearch())],
        safety_settings=[
            types.SafetySetting(
              category=types.HarmCategory.HARM_CATEGORY_HATE_SPEECH,
              threshold=types.HarmBlockThreshold.BLOCK_ONLY_HIGH,
            ),
            types.SafetySetting(
              category=types.HarmCategory.HARM_CATEGORY_HARASSMENT,
              threshold=types.HarmBlockThreshold.BLOCK_ONLY_HIGH,
            ),
            types.SafetySetting(
              category=types.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
              threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
            ),
            types.SafetySetting(
              category=types.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
              threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
            )
        ]
    )
)

Let’s assume that response was blank because of FinishReason.SAFETY.
When checking the ratings(pprint(response.candidates[0].safety_ratings, indent=2)), the reason is somehow:

[ SafetyRating(blocked=None, category=<HarmCategory.HARM_CATEGORY_HATE_SPEECH: 'HARM_CATEGORY_HATE_SPEECH'>, probability=<HarmProbability.NEGLIGIBLE: 'NEGLIGIBLE'>, probability_score=None, severity=None, severity_score=None),
  SafetyRating(blocked=None, category=<HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: 'HARM_CATEGORY_DANGEROUS_CONTENT'>, probability=<HarmProbability.NEGLIGIBLE: 'NEGLIGIBLE'>, probability_score=None, severity=None, severity_score=None),
  SafetyRating(blocked=True, category=<HarmCategory.HARM_CATEGORY_HARASSMENT: 'HARM_CATEGORY_HARASSMENT'>, probability=<HarmProbability.MEDIUM: 'MEDIUM'>, probability_score=None, severity=None, severity_score=None),
  SafetyRating(blocked=None, category=<HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: 'HARM_CATEGORY_SEXUALLY_EXPLICIT'>, probability=<HarmProbability.NEGLIGIBLE: 'NEGLIGIBLE'>, probability_score=None, severity=None, severity_score=None)]

But the block threshold is set to BLOCK_ONLY_HIGH! I doubt that this is intended. Matter of fact, it even persists with models that don’t have search grounding support at all(gemini-2.0-flash-lite). Just declare the search tool and the safety configs are broken, even if unsupported.

The behavior doesn’t persist with search tool being disabled and works vice-versa when tool is enabled(if the threshold is set to BLOCK_LOW_AND_ABOVE, it will still block according to default configuration, so medium and above).

2

Hi @reid, Thanks for reporting this. will bring this to the engineering team attention. Thank You.

1 Like