claudeIMAGE
Anthropic's sub-0.1% false-positive rate covers request blocks, not account bans
An analysis notes that Anthropic's sub-0.1% false-positive figure refers to its request classifier rather than wrongful bans of ordinary accounts, and that the article gives no account-level false-ban data.
愚公翘嘴📷
@ygqz666Archived 18 days agoOriginally posted 2026.10.13
愚公翘嘴📷 @ygqz666Originally posted 2026.10.13
Claude 的“误判率低于0.1%”，说的是请求分类器，不是普通用户账号的误封比例。 看完这条讨论再对照原文，我觉得这个口径最值得补上：Anthropic 承认正常请求也可能被拦；持续反复提取受保护推理的行为，则可能触发账号限额或封禁。 文章没有给出普通账号误封率，也不足以解释某个新号为什么被封。“请求被挡”和“账号被封”，需要分别看证据。 来源：Anthropic，2026-10-12；讨论：@op7418，2026-10-13。公司自报，非独立复测。
This post quotes another post: “Anthropic's anti-distillation post hints at how to avoid Claude bans, blogger says”
x.com/ygqz666/status/2109928377113358532Archived 2026.10.14 10:18