A report with invented sources looks exactly like a report without. Same layout, same tone, same neat footnotes. If you read it for argument and structure, as almost every reviewer does, you won't notice a thing.
That's what happened in Australia in 2025. Deloitte wrote an independent review for the Department of Employment and Workplace Relations (DEWR), covering the enforcement framework behind payments to jobseekers, under a contract worth 440,000 Australian dollars. Chris Rudge, a researcher at the University of Sydney, spotted references to work by his colleagues that didn't exist. The report also contained invented publications and a quotation from a court judgment that wasn't in the judgment. Deloitte confirmed it had used a toolchain built on Azure OpenAI GPT-4o. The department had the final instalment of the fee repaid, and the corrected version added a reference to generative AI in an appendix.
Note who found it: an outsider who happened to recognise his own field. Not the client, not a second reader.
Which phase this is
SEAL describes the four phases knowledge work goes through: Source, Edit, Approve, Launch. In every phase the traceability of a source can be lost, and the weakest phase determines what the work is worth.
An invented source is a Source problem. The question for that phase is: does the source actually exist?
The trouble is that a reviewer rarely asks it. A reference list gets read as part of the layout. You see that it's there, that it looks tidy, that the years add up. You don't open it. And a language model writes a reference that sounds so convincing, with a plausible author, a plausible title and a plausible journal, that opening it wouldn't tell you much unless you looked beyond the first line.
A sample check in three steps
You don't have to check every footnote. You do have to check the right ones. Pick the sources under the claims the recommendations rest on, plus a few at random. Ten is a good start. Then take each one through three steps.
1. Does it exist? Look up the title, the author and the year. A judgment is in the court's register, an article is with the publisher or in the journal itself. If you find nothing, that doesn't prove it doesn't exist. It is a reason to ask the author for the document. The burden of proof sits with whoever cites the source.
2. Does it say that? Open the passage and read what's there. A source can exist and still say something other than what the report claims. That happens without AI too. With a quotation, exactness is what counts: is precisely this sentence on that page?
3. Does it fit the claim? Even an existing source with the right content can carry a claim that's too broad. Look at scope, certainty and time. The article on checking an AI summary goes into that in more depth.
A made-up example to see it at work:
- In the report: "Research shows that councils that refer people early cut repeat assessments by a third (Van Dijk & Mertens, 2023)."
- Step 1: no article with this title at the publisher, no authors with this combination in the field. Ask the author for the PDF.
- Step 2: if the PDF does arrive, the "third" may turn out to be the result for a single council.
- Step 3: then "councils that refer people early" is quite a stretch.
Three steps, three different ways to be wrong.
What you ask the author
Don't ask whether AI was used. The answer is almost always yes, and it says nothing about quality. Ask what was checked, by whom, and where that's recorded.
A firm that records its sources while it works can supply, within the hour, a file with the source and passage for every claim, and the name of whoever checked it. A firm that has to reconstruct that afterwards didn't do it. Both answers are information. In TRACES, this layer is called Trail: what is recorded, and does that trail arise by itself during the work or does someone have to keep it up?
If you're the one writing the report
The reverse holds just as well. If you write a report with AI help, you hold on to the source at the moment the claim goes in, not at the final check. Then the client's sample check is a matter of opening a file.
The Deloitte report was corrected, and the department said the substance stood and the recommendations were unchanged. That may well be true. But it still cost a repayment, a correction and a newspaper story, for a check that takes an afternoon.
What you can do this week
- Take the latest report you've received or written in which AI played a part.
- Pick ten sources: six under the weightiest claims and four at random.
- Take each source through existence, content and scope, and note what you saw for each one.
- If you find even one that doesn't exist, check all of them and ask the author for the source file.
- Agree who does this for the next report, and before what point.
The number you're left with, the sources that didn't exist or said something else, is your Source leak.
Sources
- Cyber Daily (2025). Deloitte to refund government after using AI in $440,000 report. cyberdaily.au
- City AM (2025). Deloitte refunds Australian government after AI 'made up citations' in report. cityam.com