AI企业正在尝试通过多种方式教导模型拒绝危险请求。Anthropic在2021年提出了AI应该是"有帮助、诚实、无害的"这一原则1，而该公司的分类器系统增加了24%的计算成本1。然而，这些安全机制的有效性存在严重问题。意大利研究人员通过诗歌形式成功越狱了两打广泛使用的模型1，Amazon研究人员在某模型发布后不到三天内解锁了其被限制的黑客能力1，加拿大高中枪击案嫌疑人通过在问题前添加"假设地"一词欺骗ChatGPT提供了信息1。这些案例表明，模型的拒绝能力远比看起来脆弱得多。
拒绝机制的滥用风险也不容忽视。Meta的监督委员会发现，模型对涉及批评泰国国王的请求拒绝率更高，原因是泰国有不敬罪法1；OpenAI与禁止同性恋和批评政府的阿联酋建立合作1；DeepSeek R1在涉及西藏和维吾尔族的代码请求中产生了更多漏洞1。同时，拒绝能力与模型的有益能力难以分离，英国AI安全研究所发现Anthropic的模型有时会无故拒绝合理的AI安全研究任务1。这表明过度的安全限制可能被用于审查，甚至损害模型的正常功能。
Artificial intelligence companies have invested heavily in training systems designed to make models refuse harmful requests, but the reliability and broader implications of these safeguards remain deeply problematic.1 The industry's approach traces back to Anthropic's 2021 framework that AI systems should be "helpful, honest, and harmless," yet the mechanisms built to enforce refusal are proving insufficient against determined users and vulnerable to abuse.1
Early models required significant intervention to prevent indiscriminate responses—as one OpenAI safety researcher noted, earlier iterations would "blab on about anything."1 Companies now deploy specialized classifiers to block dangerous requests, though such systems exact a tangible cost; Anthropic's refusal classifier alone increases computational expense by 24 percent.1 Despite these investments, the barriers remain porous. Researchers in Italy circumvented refusal mechanisms across two dozen widely used models by embedding requests in poetic form, while Amazon researchers unlocked hacking capabilities in Fable 5 within three days of its release.1 A Canadian high school shooting suspect successfully deceived ChatGPT into providing information by framing questions hypothetically.1
Beyond technical vulnerability, AI refusal systems present a more troubling risk: they can become instruments of political censorship. Meta's oversight board documented that models declined requests criticizing Thailand's monarchy at higher rates, reflecting the country's lèse-majesté laws.1 OpenAI's partnerships with the United Arab Emirates—a nation that prohibits criticism of government and criminalizes homosexuality—underscore how safety systems may entrench authoritarian control.1 DeepSeek R1 generated more coding vulnerabilities when presented with requests involving Tibet and Uyghur-related content, suggesting geopolitical bias in refusal patterns.1 Even well-intentioned systems misfire: the UK AI Safety Institute found that Anthropic models sometimes rejected legitimate artificial intelligence safety research tasks without justification.1
评论
还没有评论，欢迎留下第一条。