In BriefSome agencies are trialling AI agents that can analyse information and coordinate tasks.Experts warn AI can shape decisions even when humans retain final approval.Hundreds of artificial intelligence (AI) use cases across 20 federal government departments have detailed experimental agents and tools being used to draft, summarise and analyse documents or workflows. But it's the lack of transparency and apparent absence of an enforcement agency to monitor or police unapproved or 'shadow' AI use that concerns experts most. Freedom of Information (FOI) releases over the past four weeks reveal AI is being used widely in government departments including the health department, ASIC, the Australian Competition and Consumer Commission, the agiculture department, the Department of Industry, Science and Resources and the Commonwealth Obudsman.The Department of Industry, Science and Resources is one department experimenting with AI agents to perform tasks. It said its trials of "agentic task automation and orchestration" are tested and evaluated, with humans checking outputs to identify and manage risks such as bias, omission or mischaracterisation of information. Among the 19 use cases, pre-built and custom AI agents conduct analysis, generate content, interact with authorised systems and "orchestrate activities across one or more data sources, tools and services".News that makes senseYour trusted source for staying up-to-date with the world around you. Get free daily news updates and analysis, straight to your inbox.Enter email addressSubscribeMorning (Mon–Fri)Afternoon (Mon–Fri)WeekendInsightBy subscribing, you agree to SBS’s terms of service and privacy policy including receiving email updates from SBS."Tools remain subject to departmental governance, defined controls, human oversight and approval by departmental staff," a spokesperson told SBS News. But Murdoch University computer science professor David Parry argues government assurances don't cater for the possibility the decision may be based on information summarised, modelled or extracted by AI tools."The self-declaration approach [by government departments] is unlikely to detect 'shadow' [unapproved] uses, or global issues across the government," he said."There doesn’t seem to be much in the way of explicit auditing or independent evaluation."The benefit of using these tools — that complex analysis can be performed with fewer or less-expert humans — leads to the difficulty of reliably assessing the outputs and detecting errors."Health Department's risky AI useIn September, auditors found unapproved or 'shadow' AI use in the federal health department, which had "not yet systematically identified, assessed and established treatments for AI risks".There were 46 novel (new) AI use cases across the health portfolio. One pilot project helped staff query and summarise Primary Health Network activity work plans. It gave links to source documents and had human oversight.But Australian Catholic University computational intelligence professor Niusha Shafiabady said attaching source links to an AI summary is a "band-aid solution" to oversight. "Health (appears to be) using generative models that paraphrase and hallucinate [make up information], and then hoping tired, overworked staff will click every link to catch the mistakes," Shafiabady said."If the goal is pulling facts from work plans, a real AI expert would engineer a purely extractive pipeline, one that retrieves verbatim information without the mathematical ability to invent anything and hallucinate. The key question is whether the system has been rigorously tested for this specific task: What is its error rate?"A health department spokesperson said the department takes a "risk-based approach" to AI, considering transparency, explainability, human oversight, privacy, risk mitigation, and accountability.  "Any use cases that do proceed to implementation must comply with the safeguards the department has in place to support the responsible use of AI and to ensure staff remain accountable for decisions made when using AI-assisted tools," they said."This includes checking outputs against source material, reviewing and collaborating with subject matter experts as required, and ensuring any use of content generated by AI remains subject to human judgement, governance controls and departmental quality standards."'Shadow AI' use by government employeesUniversity of Sydney law professor and chief investigator at the ARC Centre of Excellence on Automated Decision-Making and Society, Kimberlee Weatherall, said her recent research across Victorian government departments didn't just find shadow AI use; it was actively being surveilled by IT teams to see if it could learn how people wanted to use AI or which tools were missing. "The worry about shadow use is that things are going into [a public AI chatbot], and the extent of tracking and visibility that is occurring is terrifying," Weatherall said. This shadow AI use by government employees on personal AI accounts presents many risks, said Professor Toby Walsh, chief scientist of the AI Institute and scientia professor of AI at UNSW."I could imagine the reason that we've had all these hacking incidents into government departments is that those departments have actually been putting information into ChatGPT, which then went into the training data, which then told the agent that there's interesting information found in these government departments, and then it's gone to the government to try and find that information in more detail," he said."I'd like to know, and I don't know. Which is troubling because I’m not just a citizen, I'm actually reasonably AI qualified."Weatherall agreed, saying despite the government issuing "really quite helpful guidance and information about how to do AI well", it's guidance; it's not enforceable. "There's no significant reporting on it, the agencies effectively govern themselves and decide what they want to tell us about AI," she saidThe Digital Transformation Agency (DTA) has "never had significant enforcement power" it's purely advisory, and its main remit is to facilitate AI use, Weatherall said. "(And) this is one of the most one of the interesting things about this trove of documents. The departments have to issue these AI transparency statements and then internally they create these much more detailed use case registers that they don't have to publish or tell us about." DTA Deputy CEO Ramsey Beydoun told SBS News that 120 agencies have published AI transparency statements, detailing a high-level overview of where and how AI is used, without requiring them to publish individual use cases or use-case-level detail.New rules started in December 2025 requiring agencies to assign an accountable use case owner, mandatory training for all staff, and mechanisms to monitor and evaluate AI systems. An internal register of all AI use cases is due by 15 December 2026, but the policy does not mandate that agencies publish these."The AI Review Committee helps agencies adopt AI safely and consistently by providing expert advice on AI use cases with a residual high-risk rating and use cases that agencies determine to be highly sensitive, novel, or complex, Beydoun said. Where does AI assistance end and influence begin?At the workplace regulator, the Fair Work Ombudsman (FWO), Microsoft Copilot is used to "support first-pass review and summarisation" as well as "checking apparent continuity or gaps in pay period information".FWO staff are trained to use the technology with mandatory guardrails. All outputs are "considered strictly advisory" with "human-in-command rules" and record-keeping requirements if AI "materially informs a decision, recommendation or other action".University of Sydney Business School senior lecturer Raffaele Fabio Ciriello told SBS News that having a "human in the loop" is not a "complete safeguard"."If AI has already decided what to highlight, summarise or flag, it has already shaped the human decision," Ciriello said. "Human oversight only works when staff have enough time, expertise and information to genuinely challenge the system, otherwise the human gets reduced to a rubber stamper."The Australian government’s Microsoft Copilot evaluation didn’t assess whole-of-government cost-benefit or technically benchmark output accuracy, he said."Saving a public servant 20 minutes is evidence of task efficiency, not evidence that the public service, as a system, is better," Ciriello said.Associate investigator at the Centre for Automated Decision-Making and lecturer at QUT Law School, Henry Fraser, said using this technology for brainstorming, drafting, analysis and editing, even when it's not directly making decisions, may influence how issues are framed, what's considered, and what's left out."The extent of AI use, often undisclosed, highlights the need for binding regulation of the use of automated decision-making and machine learning technologies in government," Fraser said. "We are still awaiting draft legislation on automated decision-making in government, more than three years after the royal commission report on robodebt recommended its introduction." Southern Cross University associate law professor Brendan Walker-Munro said the claims that AI will lead to cost savings and efficiency dividends omit the cost to check and verify that the models are performing properly, safely and securely. "I think a big part of the problem is that we seem to be looking at ways to use AI to automate decisions which really need to be performed by humans," Walker-Munro said.The Assistant Minister for Science, Technology and the Digital Economy, Dr Andrew Charlton, referred questions about which agency is responsible for policing and monitoring shadow AI use in government to Public Service Minister Katy Gallagher, who then directed the questions to the Department of Home Affairs.A Home Affairs spokesperson said the department is responsible for the framework that sets security rules for federal government agencies, but did not directly answer a question relating to the surveillance of shadow AI use.It said the framework "mandates that all technology systems including AI require authorisation to operate or be used by a Commonwealth entity, ensuring that an appropriate level of security is being applied to any technology system".For the latest from SBS News, download our app and subscribe to our newsletter.Share