A groundbreaking investigation has uncovered a troubling trend sweeping through Australia’s democratic machinery. Dozens of policy submissions sent to parliamentary inquiries contain fabricated citations and invented research—hallucinations produced by artificial intelligence systems that have somehow made their way into official government documents.
Guardian Australia analysed submissions to current parliamentary inquiries and discovered at least 39 documents riddled with references to studies, journals, and academic papers that simply do not exist. The investigation used a custom-built program to extract references and cross-check them against major academic databases including CrossRef and Google Scholar. What emerged was a damning picture of how easily AI-generated misinformation can infiltrate the very heart of democratic decision-making.
The implications are stark. Parliamentarians rely on these submissions when shaping legislation and policy. If the evidence being presented is partly or entirely fictional, the decisions that follow could be built on nothing more than digital phantoms conjured by algorithms.
“When politicians are making decisions based on evidence that doesn’t exist, that’s a fundamental problem,” warns Professor Christian Downie, who leads the regulation and global governance programme at the Australian National University. His concern extends beyond poor policymaking. “We’re also starting to undermine the public’s trust and confidence in the types of institutions that underpin our democracy.”
The Hallucination Problem Explained
AI hallucinations occur when large language models—systems like ChatGPT or Claude—generate text that appears convincing but contains fabricated details. These aren’t technical glitches to be patched away. They represent an inherent limitation of how modern AI systems work: they predict the next most likely word based on statistical patterns, with no actual understanding of whether what they’re producing is true or false.
Ask an LLM to cite research supporting a particular argument, and it will happily do so—even if that research never existed. It will generate plausible author names, journal titles, publication years, and abstract summaries. The output looks professional, properly formatted, and utterly fictional.
This behaviour becomes more pronounced when AI systems tackle topics outside their training data or when users push them to generate confident responses without verification. The result is a perfect storm for anyone seeking to influence policy through written submissions, intentionally or otherwise.
When Google Becomes Part of the Problem
The situation grows even more complex when AI search tools enter the equation. Previously, fact-checking an suspicious reference might involve a quick search that returned no results—clearly indicating the source was invented. That safety net is now disappearing.
Google’s AI Overviews, the summarised answers the search giant now presents at the top of results pages, will sometimes create summaries of these fabricated references as though they were genuine. The fake citation gets a AI-generated description of its supposed contents, complete with methodology and conclusions. This makes false references appear authentic at a glance.
Even more troublingly, both Google and ChatGPT have been observed citing the very inquiry submissions containing fake references, creating a self-perpetuating cycle of misinformation. The fake source is cited by the AI summary, which then gets cited elsewhere, building a web of invented authority that becomes increasingly difficult to untangle.
One particularly alarming example appeared in a submission to an inquiry into family violence and suicide. It contained a hallucinated reference attributed to Associate Professor Divna Haslam of the University of Queensland, a respected clinical psychologist who researches child adversity and maltreatment. The submission misstated her team’s findings while attributing them to research that never took place.
Haslam describes seeing her name attached to invented work as “scary” because the fabrication was sophisticated enough to fool casual readers. “It’s very frustrating to have invested time, money, effort, expertise in rigorous research,” she told Guardian Australia. “Then to see something that’s inaccurate and inappropriately attributed anyway is really concerning.”
The company responsible, Drilldown Reports, acknowledged using AI in its research process and said it had identified the errors in a follow-up submission. However, the deadline for corrections had already passed. A company spokesperson insisted the problem was human error rather than AI, stating: “We do strongly believe in the human factor and that it should be a vital part of the full quality review process. Unfortunately, the human factor also lets us down when uploading the correct document.”
Scope of the Problem
The investigation’s findings likely represent only a fraction of the actual AI-generated material entering the parliamentary process. The methodology could only identify submissions containing references—documents that used AI to generate text without citations would escape detection entirely.
More than 100 submissions were found to contain ChatGPT URL tags in their reference links, indicating the authors had used ChatGPT to look up sources. These tags are automatically added when ChatGPT provides links to users and can persist when copied into other documents. While not definitive proof of AI misuse, the prevalence of these tags suggests widespread reliance on AI tools during the submission process.
The submissions containing fabricated references came from organisations and individuals across the political spectrum, from housing policy advocates to climate change commentators. One submission to a housing inequity inquiry appeared to cite work by Professor Nicole Gurran, an urban planning expert at the University of Sydney, that simply does not exist.
Gurran points out that citations serve a vital function in research: they connect claims to evidence, allowing others to verify and build upon previous work. “Fake citations, even if an accidental ‘collage’ where the claim is correct but the chain of reference to evidence is broken, undermines that fundamental process,” she explains.
The problem extends beyond parliamentary submissions. Last year, consulting giant Deloitte was forced to issue a partial refund to the federal government after revealing that AI tools had included fake references—and even a fake court citation—in a $440,000 report. If such a major corporation can be duped by AI-generated misinformation, what chance do smaller organisations and individual citizens have?
What Can Be Done?
Senate guidelines already warn that AI use in inquiry submissions presents risks to information quality and that accuracy remains the submitter’s responsibility. But enforcing such guidelines presents significant challenges. How do you verify that thousands of submissions from across Australia are free from AI-generated fabrications?
Professor Downie believes the solution lies in balancing openness with new safeguards. “We want our elected officials to be making decisions based on real information and real evidence, not on fake citations and fake claims,” he says. “Whether it’s climate change, immigration, health, we want that process to work properly. But we need guidelines that encourage truthfulness.”
When contacted for comment, OpenAI acknowledged that hallucinations remain an ongoing research challenge and urged users to treat ChatGPT output as “a first draft, not a final source”—emphasising that users should always verify quotes, data, and references against original documents. A Google spokesperson said AI Overviews operate “like traditional Search” in matching content to queries, suggesting the company views itself as a neutral intermediary rather than a guarantor of information accuracy.
Why it Matters
This investigation reveals a vulnerability at the core of democratic governance that demands urgent attention. Australia’s parliamentary inquiry system exists precisely to gather evidence and expertise from the public, informing decisions that affect millions of lives. When that evidence includes invented studies, fabricated research, and AI-generated misinformation, the entire system becomes compromised.
The risks extend beyond poor policy decisions. If public servants, ministers, and ordinary citizens cannot trust that government submissions contain genuine information, confidence in democratic institutions will erode further. In areas as sensitive as domestic violence, climate policy, and healthcare—where this investigation found examples of fake citations—decisions based on nonexistent evidence could cause real harm to real people.
Technology companies must accept greater responsibility for how their products are used in official processes. But governments cannot simply wait for AI firms to solve the hallucination problem; it may never be fully eliminated. Parliamentary committees need new tools, training, and verification processes to distinguish genuine research from sophisticated fabrications. Until then, the foundations of evidence-based policymaking are being quietly undermined by algorithms that don’t know the difference between truth and fiction.
