OpenAI claims matching a user's query against a policy database and sending the matching policy guidance to the model when it answers

2026-11-09Filings

Listen · Thinking Makes It So, 2026-11-09 · 9 min

All episodes · Subscribe: Apple Podcasts · Spotify

United States Patent Application Publication No. 2026/0350367 A1, OpenAI OpCo, LLC, "Conditional Inference-Time Policy Enforcement System for Generative AI Models," Application No. 19/353,131, filed October 8, 2025, claiming priority to United States Provisional Patent Application No. 63/798,123, filed May 1, 2025, published November 5, 2026

Claim 1 has four steps: access user message information for a query, compare it with a set of policies stored in a policy database, determine that it corresponds with a policy, and send that policy's guidance to the generative model for tailoring a response. Independent claim 15 adds the mechanism: the guidance goes to the model as a temporary system message, and claim 16 deletes that message from the model's thread once the model has responded; claim 7 adds modifying a policy without modifying the model's parameters. The background says hard-coded instructions in static system messages lack contextual awareness, and names voting information, medical advice and legal guidance as domains where inaccurate answers pose compliance and safety risks. The USPTO's Open Data Portal shows the application awaiting its first examination.

What it changesIn my view claim 1 as published comes close to describing retrieval of a policy text into the model's context, and the specification's own background acknowledges that instructions were already being placed in system messages. I expect claims 1 and 10 to meet prior art and the application to rest on claim 15, where the guidance arrives as a temporary system message and, in claim 16, is deleted once the model has answered. That is a sensible structure: OpenAI filed one independent claim at the broadest reading and another at the mechanism, so the medium claim is already written at the narrower scope if claim 1 falls, and claim 2, which adds the temporary system message, gives OpenAI a ready amendment for its method claim. If you are drafting guardrail or moderation claims, consider the same arrangement, with one independent claim that recites how the guidance enters the model's context, as claim 15 does, alongside the broad one.

More on What the AI companies are filing · Prior art, public disclosure and obviousness

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.