Gemini Safety Filtering

When using Gemini AI for translation or speech recognition tasks, you may encounter errors such as "Response was blocked due to safety settings."

This is because Gemini has safety restrictions on the content it processes. Although the code allows for some adjustments and the most lenient "Block None" setting is applied, the final decision to filter is still made by Gemini based on a comprehensive assessment.

Gemini API's adjustable safety filters cover the following categories. Other content not listed here cannot be adjusted through code:

Category	Description
Harassment Content	Negative or harmful comments targeting identity and/or protected attributes.
Hate Speech	Rude, disrespectful, or profane content.
Sexually Explicit	Contains references to sexual acts or other obscene content.
Dangerous Content	Promotes, facilitates, or encourages harmful acts.
Civic Integrity	Queries related to elections.

The following table describes the blocking settings that can be used in the code for each category.

For example, if you set the blocking setting for the Hate Speech category to Block Few, the system will block all parts with a high probability of containing hate speech content. However, it will allow any parts with a low probability of containing dangerous content.

Threshold (Google AI Studio)	Threshold (API)	Description
Block None	BLOCK_NONE	Display regardless of the likelihood of unsafe content.
Block Few	BLOCK_ONLY_HIGH	Block when there is a high probability of unsafe content.
Block Some	BLOCK_MEDIUM_AND_ABOVE	Block when the likelihood of unsafe content is medium or high.
Block Most	BLOCK_LOW_AND_ABOVE	Block when the likelihood of unsafe content is low, medium, or high.
N/A	HARM_BLOCK_THRESHOLD_UNSPECIFIED	Threshold not specified, use the default threshold for blocking.

The following settings can be used in the code to enable BLOCK_NONE:

safetySettings = [
    {
        "category": HarmCategory.HARM_CATEGORY_HARASSMENT,
        "threshold": HarmBlockThreshold.BLOCK_NONE,
    },
    {
        "category": HarmCategory.HARM_CATEGORY_HATE_SPEECH,
        "threshold": HarmBlockThreshold.BLOCK_NONE,
    },
    {
        "category": HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
        "threshold": HarmBlockThreshold.BLOCK_NONE,
    },
    {
        "category": HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
        "threshold": HarmBlockThreshold.BLOCK_NONE,
    },
]

model = genai.GenerativeModel('gemini-2.0-flash-exp')
model.generate_content(
                message,
                safety_settings=safetySettings
)

However, it is important to note that even if all settings are set to BLOCK_NONE, Gemini may still not allow the content and will filter it based on its inferred safety based on the context.

How to Reduce the Probability of Safety Restrictions?

Generally, the flash series has more safety restrictions, while the pro and thinking series models have relatively fewer. You can try switching to different models. In addition, when potentially sensitive content is involved, sending less content at a time and reducing the context length can also reduce the frequency of safety filtering to some extent.

How to Completely Disable Gemini's Safety Judgment and Allow All of the Above Content?

Bind a foreign credit card and switch to a paid premium account.