澳大利亚政府广泛使用AI工具但缺乏透明监管

澳大利亚联邦政府20个部门已在数百个用例中部署了AI技术，但相关使用情况鲜少向公众披露。1卫生部门在9月的审计中被发现存在46个未经批准的"影子AI"用例，这些工具被用于文件起草、总结和分析等行政工作，但部门未对其风险进行系统评估。1

政府目前正在推进新的AI治理规则，从2025年12月起要求所有员工接受强制性培训。1各部门需在2026年12月15日前提交内部AI用例登记册，但这些信息无强制性向公众公开的要求。1已有120个机构发布了AI透明度声明。1

专家对这一现状表示担忧。悉尼大学法学教授Kimberlee Weatherall指出，数字转型局从未掌握重大执法权力。1新南威尔士大学教授Toby Walsh警告称，政府部门可能直接将信息输入ChatGPT等公共工具，存在数据泄露风险。1Murdoch大学教授David Parry表示，依靠政府部门自我声明的方式不太可能发现影子AI使用。1悉尼大学商学院讲师Raffaele Fabio Ciriello强调，即使由人类保留最终决策权，AI也已通过突出、汇总或标记信息等方式影响决策过程。1

Australian federal government agencies across 20 departments are deploying artificial intelligence in hundreds of applications, yet the extent and oversight of this technology adoption remains unclear to the public 1. The use of AI tools for document drafting, summarization, and analysis has become commonplace in government operations, but transparency mechanisms and regulatory frameworks have not kept pace with this expansion 1.

In September, a health department audit uncovered 46 previously undisclosed AI applications that had been implemented without formal approval, highlighting the presence of "shadow AI" systems operating outside official channels 1. The government's Digital Transformation Agency currently functions only in an advisory capacity, lacking meaningful enforcement authority to monitor or control departmental AI deployments 1. New regulations scheduled to take effect in December 2025 will require the establishment of AI accountability officers and monitoring systems, with an internal AI registry deadline set for December 15, 2026, though there is no mandatory requirement for public disclosure of these records 1.

Experts have raised significant concerns about the governance gaps in this landscape. According to David Parry, a professor at Murdoch University, self-reporting mechanisms employed by government departments are unlikely to identify shadow AI usage 1. Kimberlee Weatherall, a law professor at the University of Sydney, emphasized that the Digital Transformation Agency has never possessed substantial enforcement powers 1. Additional risks include information security vulnerabilities, as UNSW professor Toby Walsh noted that government employees may input sensitive data into systems like ChatGPT, creating potential data leakage exposures 1. Raffaele Fabio Ciriello, a lecturer at the University of Sydney Business School, cautioned that human-in-the-loop oversight does not guarantee protection, observing that AI systems already influence decisions through prominence, summarization, and flagging of information 1. To date, 120 agencies have published AI transparency statements 1.