OpenAI has already allowed AI to take part in developing the next-generation GPT, but it has not permitted the system to recursively self-improve autonomously without human oversight. CEO Sam Altman has drawn a clear safety red line, refusing to surrender future decision-making authority to AI and placing alignment issues at the top of the agenda. OpenAI has brought in four external evaluation agencies to work on site, embedding full-process safety checks into four stages—model training, reinforcement learning training, alignment evaluation quality, and model internal mechanisms—replacing the old approach of testing only before a finished product ships. That said, external evaluators do not have independent authority to halt operations, and core internal model evaluation results are not made public, being submitted only to the OpenAI Safety Committee. At present, AI capabilities are evolving unevenly, and with the prospect that an AI self-iteration flywheel could start spinning, safety fallbacks have become an increasingly urgent core issue.
