If you’re crafting policy for a country,
the first thing you don’t usually do is ask AI:
“What do you think is the best approach?”
Instead, you start with:
research,
data,
expert opinions,
and past cases.
This is why parliamentary inquiries openly call for submissions.
Academics can submit research.
Businesses provide industry data.
Civil society groups share real-world experience.
Everyday people can tell MPs what effect a policy has on their lives.
These documents may be:
read by parliamentarians,
cited by committees,
incorporated into official reports,
and even become the basis for further policy discussion.
Underlying this is a basic assumption:
The research you present as evidence at least truly exists.
But in the AI era,
this assumption now also needs rechecking.
Australia’s Parliament Received Submissions Citing Research That Doesn’t Exist
Guardian Australia recently reviewed a large number of submissions to current Australian parliamentary inquiries.
The finding:
At least 39 submissions
contain questionable AI hallucination errors in citations.
The problem isn’t merely:
a misspelled author name.
Some citations point to real research,
but the details are mixed up —
wrong author-name,
wrong title,
wrong page numbers,
or inaccurately reported results.
In more serious cases,
the entire cited study simply does not exist.
There are even cases where AI attributed a non-existent paper to a real scholar.
This is problematic.
If fake research looks absurd,
everyone can spot it instantly.
The real danger lies in the fact that:
it looks very much like real research.
Even the Professors Named Were Momentarily Confused
One example involved Australian academic Divna Haslam.
A submission related to family violence and suicide cited a paper under her name.
The problem:
that paper does not exist,
and the submission also misrepresented the actual work by her team.
Haslam described this as unsettling.
Because the citation:
had the correct author name,
a plausible research title,
and the format resembled formal scholarly literature.
At a glance, it’s easy to believe.
Reporter and academic Margaret Simons faced a similar incident.
A submission cited an article allegedly written by her.
But she never authored it.
What’s more ridiculous is that
the fake citation was formatted so perfectly that she briefly wondered:
“Wait, did I really not write this?”
If the person whose name is falsely used can be misled by formatting,
how much time would regular readers,
research assistants,
or busy MPs scanning hundreds of pages have to verify each citation?
Guardian Didn’t Rely on AI Detectors
Here’s something very important.
Guardian didn’t just run parliamentary documents through an:
“AI detector”
and flag anything scoring “98% AI-generated.”
This is because AI detection tools are prone to errors.
Instead, they used a more straightforward, yet more reliable approach.
They extracted references from submissions,
then checked each one against
Crossref,
Google Scholar,
DOI,
and other academic databases.
If more than 20% of a submission’s citations couldn’t be found,
they proceeded with manual review
and directly contacted original authors for confirmation.
Ultimately, at least 39 submissions with suspected hallucinated citations were identified.
This number is likely conservative.
This method only catches cases where AI generated inaccurate references.
If someone used AI to write an entire document but included no references,
this approach wouldn’t detect it.
Over 100 Submissions Also Left ChatGPT URL Traces
Guardian also found that more than 100 submissions
contained reference URLs linking to ChatGPT or related AI artifacts.
But this does not prove:
“All these 100+ documents were ChatGPT-generated.”
Those markers could result from copying from
other articles,
republishing data,
or routing through different sources.
This only suggests:
large language models are likely heavily involved in document preparation.
But URL traces alone can’t determine who exactly used which AI tool.
The real confirmed issue remains:
there are indeed unverifiable studies and incorrect citations.
The Bigger Problem: Parliament Reports Then Citing These Flawed Documents
If the story ended with:
“Someone submitted a poor-quality document,”
it wouldn’t be as serious.
Parliament receives submissions daily
with varying quality,
bias,
and credibility.
What escalates the issue is that:
Guardian found some parliamentary committee reports
actually reference submissions containing numerous suspected AI hallucinations.
This means errors passed through a critical gate.
Originally, these were just AI-generated errors.
Then, they became submissions formally presented to parliament,
and finally references cited in official reports.
At this point,
the same sentence looks completely different.
If ChatGPT tells you,
“A certain study proves this,”
you might doubt it.
But if a search says,
“According to a submission received by the Australian Parliament…”
credibility suddenly increases.
The problem is:
a real official website doesn’t guarantee every citation inside it is accurate.
The Strangest Loop: AI Citing Its Own AI-Created Errors
This could be the most crucial part of the story.
Previously, if you suspected a study didn’t exist,
the simplest way was to Google it.
If nothing came up,
you doubted the citation.
Now, search engines themselves include generative AI.
Guardian found that when searching for some fake studies,
Google’s AI-generated summaries sometimes:
present these nonexistent studies as if they were real.
Why?
Because there actually exists a document online
mentioning the “study.”
That document happens to be a parliamentary inquiry submission.
The cycle goes like this:
AI invents a study that doesn’t exist.
↓
Someone includes it in an official submission.
↓
Submission gets published on a parliamentary website.
↓
Search AI finds the research cited on that official site.
↓
AI summarizes it as if it’s real.
↓
Next user sees the search result and assumes it’s legitimate research.
This creates a rarely seen phenomenon:
an information self-pollution loop.
Fake information doesn’t need to suddenly become real.
It only has to appear repeatedly in enough “reliable-looking places.”
Then it becomes more and more believable.
The Greatest Risk Is Not AI Hallucinations, But Fake Content Gaining Authority
A wrong ChatGPT answer
may only affect an individual.
A student’s incorrect citation
might impact one class.
But when errors enter:
government reports,
legal documents,
corporate research,
academic papers,
and media coverage,
the problem escalates.
Because the next users usually won’t verify from scratch.
They assume:
This is a parliamentary document.
Someone must have checked it.
This is a consultant report.
Someone’s accountable.
This is a formal research report.
Sources must exist.
AI exploits precisely this:
the trust that “someone must have verified this earlier.”
Australia Has Already Experienced a Costly Example
This isn’t the first time.
In 2025,
consultancy Deloitte produced a report for the Australian federal government containing erroneous content,
including fabricated citations and fake court cases,
and eventually agreed to a partial refund.
The government-commissioned report was worth:
AUD 440,000.
This incident reminds us:
Expensive reports,
well-known companies,
and official government contracts
do not automatically mean that every citation has been verified.
AI can make document creation extremely fast.
But it also makes:
producing “professional-looking but flawed documents” very quick and easy.
Australia’s Parliament Now Recognizes the Need for Extra Citation Checks
After Guardian’s investigation went public,
several key parliamentary committee chairs responded.
Liberal Senator Paul Scarr said
the greatest risk is decision-makers making choices based on wrong information.
His conclusion was clear:
Parliament has been warned.
Moving forward, more checks are needed on
footnotes,
sources,
and claims within documents.
But the challenge is obvious.
A major parliamentary inquiry
might receive hundreds or more documents,
each tens of pages long,
each with dozens of citations.
Verifying every paper,
DOI,
and data point manually
requires enormous manpower.
AI lowers the cost of content creation,
but raises another cost:
content verification.
The Answer Isn’t to Ban AI in Writing Submissions
MP Louise Miller-Frost reminded us of another important perspective.
AI may help people who previously struggled to participate in public policy
to better organize their views.
Not everyone has access to
a research assistant,
a lawyer,
a policy advisor,
or professional writers.
An ordinary citizen caring for a family
may have crucial frontline experience,
but not know how to put it into a formal 10-page document.
AI can assist by:
structuring,
editing,
and clearly expressing ideas.
This can increase democratic participation.
So the real question is not:
“Can AI be used in parliamentary documents?”
It’s:
“How far can AI help with writing, and what must be personally verified?”
A Simple Boundary: AI Can Organize Text, But Shouldn’t Invent Evidence
If AI helps you turn
three disorganized paragraphs into three clear ones,
the risk is relatively low.
But once AI starts saying things like:
“According to Professor So-and-so’s 2024 research…”
you should immediately pause.
Find the original paper.
Verify the author.
Verify the title.
Verify the DOI.
Check what the study actually says.
If the original source can’t be found,
don’t cite it.
This principle applies beyond parliament —
in reports,
plans,
presentations,
business analyses,
academic papers,
news articles,
SEO content,
and even day-to-day AI-assisted research work.
AI can help you organize evidence,
but don’t just assume that a nicely formatted footnote means the evidence actually exists.
The Real Challenge in the AI Era: From “Finding Answers” to “Proving Where Answers Come From”
Ten years ago,
data was scarce.
So we spent much time:
searching for information.
Now AI can deliver within seconds:
ten studies,
fifteen datasets,
five case examples.
The new challenge is:
which ones are genuine?
Do they really support this conclusion?
Are they citing another secondary source?
Where does that secondary source come from?
Ultimately, can you track back to:
firsthand evidence?
The crucial skill today might no longer be:
“Do you know the answer?”
but rather:
“Can you prove why that answer is trustworthy?”
The Most Dangerous AI Mistakes Don’t Always Look Like Mistakes
If AI tells you
“Australia’s capital is Tokyo,”
no one would believe it.
The real danger is when AI gives you
a real professor’s name,
a plausible research topic,
a polished journal citation format,
and a reasonable-sounding conclusion,
and the parliament website really cites it later.
Everything looks correct,
except one thing:
the study never existed.
So in the AI era, the biggest caution may not be obvious nonsense,
but answers that look like they’ve already been verified by someone else.
Understand AI easily every day.
sasadaily.com
Today, grow a little with AI.
Learn one AI tip daily.
Save a little time every day.
Improve a little each day.
SasaDaily grows with you.
