xAI adds a guidelines field to the Grok API for per-request moderation of NSFW, violence and profanity

Background and pain point

Teams shipping LLMs in consumer products often need different content standards for different surfaces: education, support and workplace tools want stricter output, while creative apps do not want blanket filtering. The usual fix is an extra moderation model or keyword filter in front of the LLM, which adds latency and over-blocks. On Oct 28 xAI announced on the official @grok account that the Grok API now lets every customer declare, per request, which content categories to tighten.

How it works

A request carries a guidelines array; each entry has type (always policy), a category and a behavior. Categories are nsfw, violence and profanity, and the only behavior today is decline. Grok follows the guidelines while it writes, rewriting its answer in its own words rather than blocking output afterward, and the response shape is unchanged. Unlisted categories keep the default behavior. The API validates guidelines before running the model and returns an error instead of partial output.

Benchmarks and results

xAI has not published block rates, false-positive rates or latency figures, and the docs say the details of each behavior may be refined over time. What is documented: duplicate categories return 400, unknown categories or behaviors return 422, and multi-agent models and deferred completions are rejected.

For developers

With the OpenAI Python SDK, pass guidelines in extra_body and point base_url at api.x.ai/v1; on Chat Completions it is a top-level field next to messages. It applies only to the request that carries it, so multi-turn apps must send it every turn; chat.create in the xAI Python SDK stores it. Requests with no server-side tools or files cannot combine it with stop or n greater than 1. Apps for minors or enterprises should compare outputs with and without it before dropping an external moderation layer.