Contents7
A report several hundred pages long, paid for by a government, signed by one of the largest firms in the world, with footnotes pointing to articles that do not exist. It happened twice in 2025, both times to Deloitte, which sells consulting, audit and AI services.
It would be easy to laugh. That would miss the lesson, which applies to every company whose teams write with an AI assistant: an invented citation looks exactly like a real one, and only a person who opens the link can tell the difference.
In short
In Australia and then in Canada, two Deloitte reports contained references nobody could find; these facts come from the press, which we report without having read the reports themselves.
In both cases, the firm and its client say the conclusions still hold.
The error is not specific to consulting: most people who use AI at work do not check its output.
The fix is a working rule, not a tool: open every source, read it, sign.
Two reports, two countries, the same error
The facts below are reported by third parties: they come from wire stories and press articles, not from the reports themselves, which we could not read in full. Only the publication of the Canadian plan was read at the source, on the provincial government’s website.
In Australia, according to the Associated Press story carried by Spectrum Bay News 9, a 237-page report delivered to the employment department, under a contract worth about 440,000 Australian dollars, contained an invented quote attributed to a federal judge and references to works that do not exist. A revised version was published with a disclosure that generative AI had been used. Deloitte repaid the final instalment of the contract, more than 97,000 Australian dollars according to the department, which says its recommendations are unchanged.
In Canada, the government of Newfoundland and Labrador published on 29 May 2025 a health workforce plan written by Deloitte: the needs of 21 professions over ten years, more than 100 recommendations. In November, Global News and The Canadian Press reported that an investigation by a local outlet had found citations nobody could locate. According to Canadian HR Reporter, the report runs to 526 pages, cost 1,598,485 Canadian dollars and contains at least four citations to articles that do not exist; Deloitte acknowledged the errors and stands by its conclusions.
- 29 May 2025
The province of Newfoundland and Labrador publishes the plan written by Deloitte.
- July 2025
The Australian report is published on the employment department’s website.
- 5 October 2025
The Australian press reveals the errors, the first article in a series listed by the OECD.
- 7 October 2025
The Associated Press reports the partial refund.
- 24 November 2025
The Canadian press reports the missing citations in the provincial plan.
The Australian case is listed in the OECD’s AI incidents monitor, which records AI system failures reported in the press. As of this article, we do not know whether a corrected version of the Canadian plan has been published.
Why an invented citation goes unnoticed
A language model does not look up a source: it writes the most likely continuation of a text. A well-formed footnote, with credible authors, a plausible journal name and a consistent year, is exactly what a scholarly text usually contains. So the model produces it with the same confidence as a true sentence.
How a fake reference ends up in a signed document
- The request
“back this paragraph with two studies”.
- The generation
The assistant writes two references in the expected format, one of which does not exist.
- The review
Someone reads the paragraph, nobody clicks the footnote.
- The signature
The document goes out with the company’s name on it.
The weak point is not step 2, which will always happen with some tools and some requests. It is step 3. The researcher who flagged the errors in the Australian report put it in one sentence, quoted by the Associated Press:
That’s about misstating the law to the Australian government in a report that they rely on.
This is not a consulting problem
It would be wrong to see this as a fault peculiar to consulting. The same studies that describe AI use at work show that checking is the exception. The figures below come from a survey funded by KPMG, which sells audit, consulting and AI services, and run with the University of Melbourne among more than 48,000 people in 47 countries: these are respondents, not companies.
of respondents rely on its output at work without checking its accuracy
say they have made mistakes in their work because of it
of French employees surveyed do not check the accuracy of its output
The French figure comes from KPMG France’s press release. Deloitte itself writes it in its own January 2026 study of 3,235 executives in 24 countries, in organisations it describes as on the leading edge of AI: 42% rate their AI strategy as highly prepared, and only 30% their risk management and governance (Deloitte, State of AI in the Enterprise 2026, pp. 6 and 7). The gap between the urge to move and the control of what goes out is general. The two 2025 reports simply made it visible.
Does a conclusion built on a fake source still hold?
In both cases, the firm and its client say the conclusions remain valid. That may be true. But the question deserves a straight answer: if an argument rested on a study, and the study does not exist, the argument no longer rests on anything verifiable. It may still be right, from experience or common sense, but it can no longer be presented as demonstrated.
For a business leader, the stakes are concrete. An audit report, a board memo, a tender response or a blog post commits the company’s signature. The day someone opens a footnote and finds nothing, it is not the tool that gets questioned.
Watch out
The AI disclosure came afterwards
In the Australian case, the disclosure that generative AI had been used was added in the revised version, after the errors were found. Saying from the start that a document was prepared with an AI assistant, and how it was checked, costs one line. Saying it afterwards costs trust.
Why we open every link
This blog is written with the help of AI. That is exactly why every figure it publishes is checked at the source. To prepare this series of articles, our 14 October 2026 review of the big firms’ studies recorded 489 figures from publications and the press: only 278 were kept as ready to cite, read at the source with their sample and their sponsor; the others are either to be rechecked or set aside, including 65 we could not verify and 30 that contradict each other from one source to the next. Among the discrepancies we found: a “2.5 times” return repeated everywhere, when the original document says 2.3.
We also learned it from a client, the hard way: a generated blog post described three real estate projects that did not exist. It had to be removed, and the address now answers that the page is gone for good. Since then, no article goes out without human review, and the ones on this site are born as drafts.
- Open the link.
A reference you have not opened is not a source, it is a promise.
- Find the figure on the page.
A real study can say something other than the sentence that cites it.
- Note who paid and who answered.
A firm that sells AI consulting, large companies, executives self-reporting: all of this belongs in the sentence.
- Date and sign.
The reader knows when the check was done and by whom.
This method is not technical. It takes time, and that is the time AI frees up elsewhere. We describe it for marketing content in our guide to reviewing, sourcing and signing AI-assisted content, and we apply it to reading return-on-investment studies in our analysis of AI profitability.
What it changes for a team of 20 to 500 people
A company of this size has no proofreading department. So the rule has to live in the process itself:
- Every document that goes out carries the name of a reviewer who opened the sources. Not a blanket sign-off, a person.
- AI assistants work on your documents, not on their memory: an assistant connected to a document base cites what it read, and you can trace back to the page.
- Uses are written down: what can be handed to AI, what always goes through a person. It is the basis of any approach, described in our guide to integrating AI in a business.
If you want to know where AI can save you time without putting your signature at stake, our AI Audit Sprint starts with an inventory of your actual uses, including the ones nobody has declared. And so that the approved tool costs what it should, our AI at the right cost offer prices each use case.
Frequently asked questions
Did Deloitte acknowledge using AI?
For the Australian report, the revised version discloses the use of generative AI, according to the Associated Press. For the Canadian plan, the press reports that Deloitte acknowledged the incorrect citations and stood by its conclusions. These facts come from the press, not from reading the reports.
Why does an AI invent references?
Because a language model writes the most likely continuation of a text, without consulting a library. A well-formed reference is likely in a scholarly text, whether it exists or not. An assistant connected to real documents reduces the risk without removing it.
How do you check the sources of a document written with AI?
Open every link, find the cited figure on the original page, note who commissioned the study and who answered, then date the check. A reference without a link is searched by its exact title; if it cannot be found, it comes out of the document.
Should AI be banned for writing reports?
No. A ban pushes use into the shadows without removing the risk. It is better to write down who reviews, what must be checked, and to put on each document the name of the person who opened the sources. The time saved remains, the risk is under control.
Sources
- The Associated Press story carried by Spectrum Bay News 9 (new tab)baynews9.com
- Published on 29 May 2025 (new tab)gov.nl.ca
- Global News and The Canadian Press (new tab)globalnews.ca
- Canadian HR Reporter (new tab)hrreporter.com
- OECD’s AI incidents monitor (new tab)oecd.ai
- KPMG France’s press release (new tab)kpmg.com
- Deloitte, State of AI in the Enterprise 2026, pp. 6 and 7 (new tab)deloitte.com
Jim · founder of Brain Plus AI
Also in business ai
Practical guide8 min
AI-written content: review, source, sign
An AI model writes fast, but what goes out in the company's name commits it. Three steps make a generated text publishable: a review that checks facts and promises, a source for every number, and a person who signs.
Stepping back11 min
AI return on investment: why it is taking so long
The figure of 95% of AI projects with no return went round the world. What it really measures, why the other studies say roughly the same thing, and the practical lessons to draw from them for an SME.
Practical guide11 min The topic guide
Integrating AI in business: private and at the right cost
For teams of 20 to 500 people, integrating AI comes down to answering five questions in order: which task, where the data goes, which model, at what price, under what rules. This guide takes them one by one and points to our detailed articles.