How does safety filtering architecture work in ChatGPT pipelines?

Safety filtering uses pre-processing, model-level alignment, and post-processing checks to ensure safe outputs.

ChatGPT safety architecture operates in multiple layers. First, input moderation filters detect harmful or disallowed content before it reaches the model. Second, the model itself is aligned using RLHF to reduce unsafe outputs. Third, output filtering checks generated responses before they are returned to users.

Some systems also include policy models that classify prompts and responses for risk. This layered approach reduces the probability of unsafe or policy-violating outputs while maintaining usability.