Record·USPTO
20260161958 OpenAI SafetyPatent
The patent application discloses methods for training and operating language models that reason over learned policies to choose compliant responses, refusals, or modified completions.
US 2026/0161958 A1doc date 2026-06-11source: PDF