aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score53

Anthropic has Claude assess misuse risk after each responseMachine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI introduction

After responding, Claude now makes a preliminary safety assessment of the conversation and sends its judgment to Anthropic’s reviewers. Anthropic says the system has surfaced misuse that its other safeguards were not designed to catch since it began using it in June. On Fable 5 traffic, the assessment matched 78% of the original detector’s flags and 94% of its high-priority flags, with a 1% false-alarm rate.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

每次 Claude 作答后,现在都会悄悄自问是否刚刚帮助某人造成了伤害,而这个私下作出的判断会交给 Anthropic 的审核人员。

Anthropic 最准确的滥用检测器是一个 Claude Mythos Preview 分类器,它在每次作出判断前都会读取其全部安全政策,因此用于每次聊天的成本过高。

因此,Anthropic 训练 Claude 在不读取这些政策、也不进行相应推理的情况下模仿该检测器的判断;在 Fable 5 的流量中,这一模仿版本与检测器的标记有 78% 一致,对高优先级案例的一致率升至 94%,误报率为 1%。

QuoteAnthropic@AnthropicAI
Translation
我们已训练 Claude 在对话中标记潜在的滥用行为,我们称这种技术为自我评估。Claude 在作出回复后,会立即对对话进行初步安全评估。 自从我们在 6 月开始使用自我评估以来,它发现了我们其他防护措施并非旨在捕捉的滥用行为。 由于 Claude 已经读过对话,自我评估的效率很高。而且,与针对一组特定危害的分类器不同,它依靠 Claude 自身的判断,因此随着我们的模型改进,它也会变得更加准确。 在我们新推出的 Safeguards 博客上阅读更多内容: https://t.co/18j6zdHub4
Original

We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation. Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch. Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve. Read more on our new Safeguards blog: https://t.co/18j6zdHub4

View this post on X ↗

Source: Rohan Paul · x.com

Products
Found an error? Send a correction