An Australian parliamentary submission cites a paper that the named academic never wrote. Once the submission is published on a government website, search systems can find it and AI summaries may describe the paper as real.
A Guardian Australia investigation found at least 39 inquiry submissions with apparently fabricated or incorrect references. Some contained a few bad citations. In others, every reference checked by the reporters appeared not to exist.
How the reporters checked the references
For the investigation, the Guardian built a custom tool that extracted references and checked them against Crossref, Google Scholar and DOI records. Reporters manually reviewed documents when at least 20% of their references could not be matched. They also contacted many of the named authors.
More than 100 papers contained ChatGPT URL tags. The Guardian treated this as a limited signal because the tags can remain when someone copies a link from another source. They do not prove direct ChatGPT use.
A reference may look polished even when the paper was never published.
How one bad citation gains authority
An AI-assisted draft introduces a false reference
A submission cites a realistic title, author or journal, but the paper cannot be found.
The submission is published
The document enters the public record on a parliamentary website, where search systems can reach it.
Search and AI systems find the page
The false reference now sits on a genuine government domain and can be retrieved in a later search.
The error returns as an AI-generated answer
An AI summary describes the nonexistent study and cites the submission as support.
The official page feeds the next answerA government URL can still contain an unverified claim
The Australian Parliament website is an authentic public record, but many submissions come from outside individuals and organisations. Their publication does not verify every claim or citation.
A search system may find the government page and repeat its claims without recognising the document as an outside submission or checking whether the cited paper exists.
Real names attached to invented papers
One submission cited a nonexistent work attributed to University of Queensland researcher Divna Haslam and misstated her team's findings. Google's AI summary later described the reference as a genuine paper.
Another submission cited work that University of Sydney professor Nicole Gurran had never published. Journalist and academic Margaret Simons was credited with a journal article she had not written. Its formatting was convincing enough that Simons briefly questioned whether the article existed.
What the investigation established
False or mismatched references reached parliamentary inquiry systems, and some committee reports cited submissions that contained apparently hallucinated material. The investigation did not show that a specific invented paper changed Australian law.
Drilldown Reports acknowledged using AI in its research process. It said a follow-up submission identified the errors, but a human upload mistake left the wrong document online after the correction deadline.
The submitter remains responsible
The Australian Parliament's submission guidance warns that AI can affect quality and accuracy. Submitters remain responsible for all information they provide, including AI-generated material.
Before submitting a document:
- Search every paper title and author combination.
- Resolve every DOI and compare its metadata with the citation.
- Open the original work and confirm that it supports the attached claim.
- Remove a reference if the underlying publication cannot be found.
Where AI FactScan fits
AI FactScan checks links in AI answers and flags broken destinations, search-result pages, self-references and weaker source types.
Unlinked bibliography entries still need to be checked in academic databases. Confirming that a paper supports a claim also requires reading the original work.
Early checks matter. It is easier to remove a bad citation from an AI answer than from a report already published and indexed online.
AI FactScan