OpenAIPatent applicationAI & softwareInterest 8/10

OpenAI patents safety checks that run while AI speaks

Illustration of the invention
Illustration based on the patent drawing. Not a photo of a real product.
Patent drawing
Patent drawing from the official filing.

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.

Why it mattersThis approach could make AI assistants feel more responsive by hiding the delay of safety checks, which is a key challenge for real-time voice interactions.

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 & computing

More from OpenAI in AI & software

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.