Guardian Australia found at least 39 submissions to Australian parliamentary inquiries containing references that appear to be fabricated, and more than 100 further documents carrying ChatGPT tracking tags inside their links. Submissions are the mechanism by which parliament takes evidence from experts and citizens before it legislates. In some cases, parliamentary reports went on to cite documents in which most of the sources, on checking, turned out to be AI hallucinations.
The most consequential part of the finding is not the count. It is the loop. Searching for a fabricated citation used to expose it: no such paper, no result. Google's AI summaries now sometimes retell a non-existent source as though it were real. And in some instances, Google's summary and ChatGPT cite the parliamentary submission that contains the fake reference in the first place. The fabrication acquires a government document as its provenance, and the search tools launder it back out as fact.
One case shows what that costs a working researcher. A submission to an inquiry into family violence and suicide carried an invented reference attributed to Divna Haslam, an associate professor at the University of Queensland and a clinical psychologist, and misstated the findings of her team's research. Haslam studies adverse childhood and family experience and maltreatment. She said the reference looked plausible enough that anyone skimming the document would take it for real, and that Google's AI summary retold it as a genuine paper. She spent time, money, effort and professional expertise producing careful research, she said, and then saw inaccurate information attributed to her regardless. She called the situation alarming.
Drilldown Reports, which prepared the document, told Guardian Australia it had used AI in its research process and found errors in a later version of the submission, by which time the correction window had closed. A representative said full quality control requires a human in the loop, and acknowledged that the human factor had failed the company and that a mistake was made in uploading the correct document. Drilldown Reports said this was human error, not AI error.
The chair of the Standing Committee on Social Policy and Legal Affairs, which oversaw the inquiry, said committees receive material of varying quality and from varying positions, and that the committee's job is to accept the evidence submitted and then test it during the inquiry.
Haslam's view is that cases like this devalue real research and create serious risk if misleading information reaches public policy, with domestic violence a particularly dangerous field for it.
The method behind the count is worth stating, because it sets the floor. Guardian journalists built a program that extracted the references from every submission filed to parliamentary inquiries during the current parliament and matched them against academic databases — CrossRef, which indexes papers and books, and Google Scholar. Where a document supplied digital object identifiers, the program also checked whether the DOI resolved to a reachable URL. Any document where 20% or more of the references failed to match went to manual review, and a large set of documents with bad references was then checked against their named authors to confirm the method worked.
The reporters deliberately avoided commercial AI detectors, which produce false positives that can see a person wrongly accused of using ChatGPT, Claude or another service. The substitute signal is citation damage: analysis of AI-written text and of AI-written scientific papers shows the models routinely mangle references, getting pages, titles and author names wrong or inventing the details outright. The links were separately checked for ChatGPT tags, which the platform attaches automatically to URLs it suggests and which survive a copy-paste into a document — though a tag alone does not prove the author used ChatGPT directly, since the text may have been lifted from a third-party article.
All of which means 39 is a conservative number. The method only catches documents that have citations at all. A submission written by a model with no references attached leaves no trace, so the true extent of AI use is certainly higher.
Two more cases show the pattern is not confined to one policy area. In an inquiry into housing market inequality this year, references appeared to attribute non-existent work to Nicole Gurran, a professor of urban and regional planning at the University of Sydney. Gurran noted that transparency and verifiability are core to research and peer review, which is why citations exist as evidence for claims — and that fake references undermine that principle even when the underlying claim happens to be true and only the chain of supporting sources is wrong. A separate submission to Senate hearings on climate misinformation last year, from an organisation called National Rational Energy Network, described the Guardian as a left-leaning newspaper linked to activists and appeared to attribute part of that claim to an article by the journalist and academic Margaret Simons in a leading scientific journal. Simons, who sits on the board of the Scott Trust, the Guardian's owner, never wrote it. The endnote so exactly matched the formatting of the other references that she briefly wondered whether the paper existed, despite knowing she had not written it. National Rational Energy Network did not respond to a request for comment.
The committee chair's answer is the place to look hardest. Accept the evidence, then test it during the inquiry is a reasonable description of how a parliamentary committee has always worked — and it describes a verification stage calibrated for human error rates. A person who fabricates a citation has to decide to do it, and can only do it so many times. A model produces a plausible endnote in the same keystroke as the sentence it supports, at no cost, in unlimited quantity, and, as Simons found, in formatting indistinguishable from the real ones. Senate guidance for submitters says AI use can affect the quality of information and that accuracy is the submitter's responsibility. That rule assumes a submitter who knows what they filed. Guardian Australia found that many of the people it contacted did not know models could fail this way, and others understood the risk but missed the errors before sending. The documents ranged from a few bad citations to submissions in which not a single reference existed, from individuals and organisations across the political spectrum.
What nobody in this account addresses is the remediation. Nothing here says whether the 39 submissions have been withdrawn or corrected, whether the parliamentary reports that cited them are being re-examined, or what consequence attaches to filing an invented citation with a committee — and the last of those is the one that would change behaviour. Last year Deloitte partially refunded the federal government after AI tools added fabricated references, including a non-existent court case, to a report costing $440,000. That is the only instance here in which anything came back.
The vendors were asked and answered past the question. OpenAI said eliminating hallucinations remains an area of ongoing research and advised users to treat ChatGPT as a draft rather than a final source, verifying quotes, data and links themselves. Google said its AI summaries work like traditional search, matching content on the web against the words in a query, and that the summaries surface web pages containing those words in the same way ordinary blue links do. Google was asked how it limits its role in spreading bad information and replied with a description of how retrieval works. It is a non-answer to the specific failure documented here, which is a summary presenting a source that does not exist as one that does.
Christian Downie, professor at the Australian National University's School of Regulation and Global Governance, framed the stakes beyond any single bad decision: if this persists, parliamentarians risk deciding on the basis of evidence that does not exist, and the discovery of fake material or citations in government submissions and even court judgments erodes public trust in the institutions democracy runs on. He argued the inquiry process should stay as open as possible, while new rules may be needed to reward truthfulness — on climate, immigration and health, elected representatives should be deciding from real information rather than fake references and invented claims.
Misleading claims and invented citations predate all of this. What changed is throughput. The submission process is open by design, and its openness was affordable when writing a plausible false document took a person days of work. The cost of filing has collapsed; the cost of checking has not moved at all, and it still falls on committee staff, on unpaid reviewers, and on researchers like Haslam who discover their names attached to work they never did.