OpenAI patents safety checks that run while AI speaks


OpenAI describes a system that checks AI-generated text and audio for safety issues as they are produced, rather than waiting for the full response to finish. It uses a small buffer to let a few words pass through immediately so the conversation feels natural, while the rest are verified in the background.
In the filing’s own words
The present technology pertains to optimizing the performance and safety of generative response engines used to produce multi-modal outputs, including audio, text, and visual media. The present technology introduces guardrails and safety checks without degrading user experience through perceptible lag, particularly in conversational or speech modes. The technology employs tunable processing to ensure compliance with safety guidelines while maintaining real-time output. It utilizes a main queue to index output tokens, feeding them into both a safety queue and an output queue. The system allows simultaneous token processing; the safety engine performs checks while a tunable buffer enables a limited number of tokens to bypass immediate safety verification to maintain conversational flow.
What exactly they claim (claim 1)
1. A method comprising: receiving, at a safety engine, a stream of tokens, wherein the stream of tokens is output by a generative response engine and wherein respective tokens in the stream of tokens are assigned an index; processing, by the safety engine, a respective token from the stream of tokens using a safety plugin; and outputting, to a front end of the generative response engine, tokens from the token stream while a difference between an index of a respective output token and an index of the respective token processed by the safety plugin is less than a buffer.
Facts
- Company
- OpenAI (OpenAI OpCo, LLC.)
- Official title
- SAFETY ENGINE PROCESSING OF GENERATIVE RESPONSE ENGINE OUTPUT
- Publication number
- US20260340286A1
- Application number
- 19189837
- Filed
- Apr 25, 2025
- Published
- Oct 29, 2026
- Inventors
- 9
- Claims
- 20
Classified under
G06F 40/284 Physics & computingMore from OpenAI in AI & software
- OpenAI patents voice system that detects when you stop talking Sep 22, 2026
- OpenAI patents remote control for AI agents running on your device Nov 3, 2026
- OpenAI patents system to catch AI models cheating on tasks Sep 10, 2026
- OpenAI patents method to check AI thinking for cheating Sep 10, 2026
- OpenAI patents system that turns photos into new AI-generated images Sep 10, 2026
Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.