What does “Safety Filter” mean?
A moderation layer that blocks or rewrites unsafe model inputs or outputs.
How Safety Filter Works
User input
Input filter
Model
Output filter
Delivered reply, or blocked.
Practical Example
A chatbot for kids has a layer that checks each reply. If a draft contains disallowed content, the filter blocks it or replaces it with a safe message before it appears.
Why It Matters
Filters add a defense independent of the model's own training. They can wrongly block harmless content or miss clever evasions.
Common Questions About Safety Filter
Q: What do filters catch?
Categories such as hate, self-harm, explicit content, or personal data, depending on how they are set up.
Q: Are they perfect?
No. They produce both false positives and false negatives.
Q: Are filters built by classifiers?
Often, a smaller model or rule set scores text for risk categories.