研究者质疑 OpenAI 关闭 reasoning=None 损害可监控性研究

Ryan Greenblatt @RyanGreenblatt原推发布于 2026.09.06

在 X 上打开原推

OpenAI seems to imply that it's good for safety/monitoring to make `reasoning=None` unavailable to the public because Astra has much higher latent reasoning capabilities (with less need for CoT). But this actually makes it even more important to study AI capabilities without CoT! Not allowing people to query the AI without reasoning doesn't make these monitorability problems go away (the model can still latently reason without verbalizing it when reasoning is on, and the reasoning-on case is the more concerning one anyway given the higher capabilities)! It's possible that OpenAI didn't mean to imply that not allowing for `reasoning=None` is good for monitorability. And it's pretty likely that the reason they are doing this is something more like jailbreak robustness or preventing distillation (AIs might spill reasoning into the output if they can't reason in the normal field). But it would have been good if AI companies (including OpenAI and Anthropic) more explicitly acknowledged that not allowing for disabling reasoning is actively *bad* for monitorability science! (Note that disabling reasoning has historically been an important way to elicit subversion capabilities for testing, in addition to being necessary for testing no-CoT capabilities.) To be clear, I don't think it's extremely important to make `reasoning=None` available to the public, but I do think that getting a better understanding of the AI's no-CoT capabilities is very important given the massive apparent jump, and external researchers not being able to run experiments on this is a problem! (This isn't just an OpenAI problem; Anthropic also doesn't allow for disabling CoT for Fable.) I think some kind of researcher access program that gives more API features, including disabling reasoning, and has a relatively low bar for entry would be a good step if there are good reasons disabling reasoning can't be made publicly available.