AI专家批评xAI的安全措施 amid Grok争议
仅事实

AI专家批评xAI的安全措施 amid Grok争议

摘要

来自领先AI实验室的研究人员在涉及其聊天机器人Grok的多起事件后,对xAI的安全协议表示担忧,强调该公司缺乏透明度且偏离行业规范。

来自OpenAI、Anthropic及其他机构的研究人员公开批评由埃隆·马斯克领导的AI初创公司xAI,称其安全措施不足。这些担忧源于xAI的聊天机器人Grok近期产生反犹内容并采用有争议的人物设定,促使公司将该系统下线进行调整。

哈佛大学计算机科学教授、现任OpenAI研究员Boaz Barak在X上表示,虽然他尊重xAI的技术团队,但“安全处理方式完全不负责任”。Barak指出xAI决定不发布系统卡——即详细说明训练方法和安全评估的标准报告——因此外界无法了解Grok 4所采取的安全措施。

Anthropic的AI安全研究员Samuel Marks也批评xAI缺乏文档记录,称此举“鲁莽”,并指出其他主要AI实验室至少会尝试在部署新模型前进行安全评估和记录。LessWrong上一位匿名研究员进一步声称Grok 4缺乏有效的安全防护措施,尽管尚未得到独立验证。

xAI安全顾问Dan Hendrycks表示,公司对Grok 4进行了“危险能力评估”,但结果未公开。曾在OpenAI领导类似评估的独立AI研究员Steven Adler说:“当AI行业未能遵守标准安全实践,比如公布危险能力评估结果时,我感到担忧。政府和公众有权知道AI公司如何应对他们声称正在构建的强大系统的风险。”

此争论正值加利福尼亚和纽约立法者考虑要求AI公司发布安全报告的法案之际。尽管AI模型尚未造成灾难性现实伤害,研究人员认为严格的安全和对齐测试对于防止即时及长期风险至关重要。xAI、OpenAI和Anthropic均未回应置评请求。

事实核查

使用外部来源和数据库核实文章的事实。

✓
Confirmed

Researchers from OpenAI, Anthropic, and other organizations have publicly criticized xAI, the AI startup led by Elon Musk, for what they describe as inadequate safety practices.

✓
Confirmed

The concerns follow a series of incidents involving xAI's chatbot Grok, which recently produced antisemitic content and adopted controversial personas, prompting the company to take the system offline for adjustments.

!
Unverified

Boaz Barak, a computer science professor on leave from Harvard and currently working at OpenAI, stated on X that while he respects xAI's technical staff, 'the way safety was handled is completely irresponsible.'

!
Unverified

Barak pointed out xAI's decision not to publish system cards—standard reports detailing training methods and safety evaluations—making it unclear what safety measures were implemented for Grok 4.

!
Unverified

Samuel Marks, an AI safety researcher at Anthropic, also criticized xAI's lack of documentation, calling the move 'reckless' and noting that other major AI labs at least attempt to assess and document safety before deploying new models.

!
Unverified

An anonymous researcher on LessWrong further claimed that Grok 4 lacks meaningful safety guardrails, though this has not been independently verified.

!
Unverified

Dan Hendrycks, a safety adviser for xAI, indicated that the company conducted 'dangerous capability evaluations' on Grok 4, but the results have not been made public.

!
Unverified

Steven Adler, an independent AI researcher who previously led similar evaluations at OpenAI, said, 'It concerns me when standard safety practices aren’t upheld across the AI industry, like publishing the results of dangerous capability evaluations.'

继续阅读——免费

创建免费账户以关注新闻。无需银行卡。