The Australian government's contract register carries one line that most of the coverage of this story never quoted. Contract notice CN4118426, Department of Employment and Workplace Relations, supplier Deloitte Touche Tohmatsu, value A$439,142.00, description: "Assurance Review to support the Targeted Compliance Framework. Final Instalment of $97,587.11 (GST inclusive) repaid to department."
Divide the second number by the first and you get 22.2222 per cent. The match is too clean for a negotiated figure. It is two-ninths, to the cent: 439,142.00 multiplied by two and divided by nine is 97,587.11. The refund was the last instalment on a payment schedule, and the schedule was written in ninths before anyone knew the report would need correcting.
So the headline number is a schedule fraction, and the honest version of this essay has to start there. Nobody sat down and valued the fabricated citations at A$97,587.11. What happened is narrower and, I think, more useful: a buyer looked at one document, decided the analysis in it was worth keeping, decided the sourcing in it was not, and settled on the one payment it had not yet made. The price of the sourcing became whatever the schedule said was left. That is still a price. It is the first one I can find where a client and a vendor agreed, in public, on what the provenance of a document was worth separately from the document.
What the buyer said, in the buyer's words
The department's statement of 3 October 2025, over the Secretary's name, spends four sentences on the matter, under the heading "Independent assurance review":
There have been media reports indicating concerns about citation accuracies which were contained in these reports. Deloitte conducted this independent assurance review and has confirmed some footnotes and references were incorrect. A correct version of the statement of assurance and final report has been released. The department continues to focus efforts on addressing the substance and recommendations included in the report.
The statement does not mention money. The refund exists in the department's own record only as that description line on the contract notice, and in what the department told reporters. The Associated Press, on 7 October 2025, quoted Deloitte saying "the matter has been resolved directly with the client" and reported that Deloitte declined to say whether the errors were generated by AI.
Deloitte's own account is inside the deliverable. Page 2 of the corrected report, under "Report Update", reads:
This Report was updated on 26 September 2025 and replaces the Report dated 4 July 2025. The Report has been updated to correct those citations and reference list entries which contained errors in the previously issued version, to amend the summary of the Amato proceeding which contained errors, and to make revisions to improve clarity and readability. The updates made in no way impact or affect the substantive content, findings and recommendations in the Report.
Read those two passages together and the separation is explicit on both sides. The department says it is working on "the substance and recommendations". Deloitte says the corrections "in no way impact or affect the substantive content, findings and recommendations". Two parties, one document, and both of them drew the same line through it: the analysis on one side, the citations and one case summary on the other. The money moved along that line.
What was wrong, counted from the bytes
The department publishes the report at one download address, and the Internet Archive captured that address three times, so all three versions can be read side by side. I extracted the text of each and counted.
The 4 July 2025 version runs to 234 pages. It cites a book called The Rule of Law and Administrative Justice in the Welfare State: A Study of Centrelink, attributed to Lisa Burton Crawford and Federation Press, 2021, nine times across seven pages: three times in footnotes on pages 58, 73 and 75, with page pins like "112–15" and "45–6", and six times in the reference list on pages 207 to 210, as entries 48, 59, 82, 89, 101 and 115. Christopher Rudge, deputy director of Sydney Health Law at the University of Sydney, was the person who noticed the book does not exist; the Australian Financial Review carried his finding on 22 and 25 August 2025, and the department's statement dates the report's publication to 14 August. In the September version the title appears zero times.
The other headline error is on page 58 of the July version, in a boxed summary of Amato v Commonwealth, the Robodebt test case. The footnote cites "Amato v Commonwealth of Australia [2021] FCA 1019, [47] – 50". The box then says "her Honour Justice Davis stated at [25] -[26]" and quotes two sentences, then "Her Honour further asserted in her ruling at [30]" and quotes two more. The September version replaces the whole box. The new heading reads "Deanna Amato v Commonwealth of Australia, Federal Court of Australia, VID611/2019, 27 Nov 2019", the text says the proceedings "were resolved by way of consent orders", and the only quotation left is attributed to "the notes to the consent orders, Her Honour at paragraph 9". The 2021 citation, the misspelt judge, the pinpoints at [25], [26] and [30] and the words attributed to them are gone. Deloitte's page 2 calls the original summary one that "contained errors". I have not read the court file, so I will not go further than the two versions and that sentence do. Rudge, quoted by The Nightly on 6 October, went further: the quotation was made up.
And then the disclosure. The July version does not contain the string "GPT" or "Azure" on any page. The September version contains them on two pages. Page 48, in the methodology, lists among the technical workstream's tasks a traceability assessment "which included the use of a generative AI large language model (Azure OpenAI GPT-4o) based tool chain". Page 147, in the appendix on method, repeats it: "a generative artificial intelligence (AI) large language model (Azure OpenAI GPT-4o) based tool chain licensed by DEWR and hosted on DEWR's Azure tenancy." Both sentences scope the tool to one job, assessing "whether system code state could be mapped to business requirements and compliance needs". Neither says the model drafted prose or produced citations. Nobody on the record has said that. What the record supports is that a model was in the toolchain of the review, that the disclosure of it arrived only in the corrected version, and that the vendor would not answer the question of where the fabrications came from.
Why the price is a price anyway
A schedule fraction is a weak valuation, and I have said so. Here is why I still think the number means something.
The department did not have to settle for the instalment. The Greens, at Senate estimates in the week of 8 October 2025, asked for the whole A$440,000 back, and the department's Secretary, Natalie James, told the hearing, as PS News reported it, that "we should not be receiving work that has glaring errors in footnotes and sources". The same report says the department's letter asking for the money named the figure, A$97,587.11, and gave two grounds: "the issues identified with the final product" and Deloitte's "acknowledgement that internal Deloitte policy regarding communicating the use of AI to the client was not followed". I could not open the letter itself; the Hansard of the hearing and a Finance department FOI release that appears to hold it both refused every fetch I tried, so those quotations are reported speech and I am marking them as such.
But the shape is clear from the documents I could open. A buyer in a position to demand more, with a parliamentary committee pressing it to, chose to keep the analysis and take back one instalment. The report kept its findings and its 200-odd pages of business-rule assessment. The buyer's stated reason for the refund was the sourcing and the undisclosed tool, which are the two components of provenance: where the claims came from, and how the document was made. Two-ninths of the fee is what the buyer decided those were worth, given that it was keeping the rest. That the number came off a schedule rather than out of a valuation is how most prices in procurement get set. It is still the number both parties signed off on, and it is on the public register.
Courts have fined lawyers for fabricated citations, and those figures get quoted as the cost of hallucination. They are penalties set by a judge against a professional. This is different in kind: a price agreed between a buyer and a seller, with the buyer keeping the goods.
The correction was corrected, and the corrections were in the correction
The department's page for the report now carries a version note nobody reported: "This Report was updated on 3 February 2026 to address identified corrections. It replaces the report dated 26 September 2025 previously published on 3 October 2025." The page's metadata gives the modified date as 3 February 2026.
I diffed the September text against the February text. The February version is 237 pages, the same as September. Its extracted text is three characters longer. Six pages differ: 12, 26, 27, 59, 77 and 78. Page 78 changes only in trailing spaces. On the other five pages the changes are to footnotes 14, 21, 24, 58, 59, 60 and 96, and they are of two kinds.
The first kind is a fix to an error the September correction introduced. The July report cited Ian Ayres and John Braithwaite's Responsive Regulation: Transcending the Deregulation Debate (Oxford University Press, 1992) correctly, and spelt Ayres correctly on every page: the string "Ayers" appears zero times in July. The September version spells it "Ayers" in four footnotes, and in two footnotes, 14 and 96, it cites the book as "John Braithwaite, Responsive Regulation: Transcending the Deregulation Debate (Oxford University Press, 2002)", which drops the first author and gives the 1992 book the year of a different Braithwaite book. The February version restores "Ian Ayres and John Braithwaite" and "1992" in both places and fixes the spelling in all four.
The second kind is spacing inside URLs and author names, of the sort that survives copy and paste between documents.
That is the whole February correction, as far as text extraction can see it. The cover of the February file still says "Final Report 26 September 2025", and the page 2 note still describes only the September update. A reader who downloads the current report has no way to learn from the document that it is the third version.
This matters for the thesis, and it cuts against the tidy form of it. If the analysis and the sourcing were separable the way the price implies, the sourcing should have been fixed once and stayed fixed. Instead the September repair, done under public pressure and priced at two-ninths of the fee, mis-cited a book it had cited correctly in July and misspelt an author on four pages, and the buyer needed a third version four months after the money moved. The provenance channel kept costing effort after it had been priced. Whether Deloitte or the department did the February work, and what it cost, I do not know; the register shows no amendment beyond the one description line.
What the buyer bought, by channel
A deliverable like this one carries three things. There is the analysis: 237 pages of findings about whether 370-odd business rules in a welfare compliance system match the law that authorises them. There is the sourcing: the footnotes and the reference list, which are the report's claim that its analysis rests on something outside itself. And there is the record of how it was made, which in this case was a two-sentence disclosure added in September at pages 48 and 147.
The department paid in full for the first. It got two-ninths of the fee back for the second, which is the fraction the schedule left. The third arrived last and was never priced at all; it was added to the corrected version and cost the buyer nothing beyond having to ask for it, which, on the reported account of the department's letter, it did.
I said at the top that this is the first public market price on provenance that I can find. I want to be exact about the claim. I searched the coverage of this case and the sanctions cases that get cited alongside it; I did not run a systematic search of procurement registers, and I would expect quiet settlements of this shape to exist that never reached a register. This one did, at CN4118426, in a description field, in a sentence nobody had to write. That is why the number is worth more than the round A$97,000 the wire stories carried. The cents are on the register, and the cents are what tell you it was a schedule.
Sourcing notes: every figure in this piece comes from a document I opened, and the two that I computed come from scripts named below. The contract value and the refund are the AusTender contract notice's own fields, read on 5 September 2026. The three versions of the report are the department's own download address as captured by the Internet Archive on 16 August 2025 (the July file, 234 pages), 6 October 2025 (the September file, 237 pages) and 28 July 2026 (the current file, 237 pages); the page numbers above are the PDF page numbers of those files. The department's 3 October 2025 statement and the report page's version note were read from Internet Archive captures of dewr.gov.au taken on 3 October 2025 and 22 August 2026, because the live site refused every fetch from my connection during the writing. The Senate estimates material is reported speech from PS News and the Australian Greens; the Hansard and the Finance department's FOI document were not opened. Rudge's finding is credited to him from the Australian Financial Review's dates and The Nightly's interview; the counts of the fabricated citations are my own, from the July file.
Reproduction
refund_arithmetic.py: the share, the retained amount and the two-ninths check, from the two AusTender figures. Output: share refunded 22.2222 per cent; kept A$341,554.89; "exactly 2/9 of the contract value: 2/9 x 439,142.00 = 97587.11".diff_report_versions.py: page and character counts of the three versions; occurrences of the fabricated title (9 on 7 pages, then 0, then 0), "Davis" (1, 0, 0), "Ayers" (0, 4, 0) and "GPT" (0, 2, 2); the July-to-September change (2,262 lines removed or changed, 2,198 added or changed, 231 of 234 pages differing, most of it renumbered footnotes); and the September-to-February change in full (25 lines each way on pages 12, 26, 27, 59, 77 and 78). Text is extracted with pypdf, so anything drawn as an image is outside the count.The files:
PRIMARIES/austender_CN4118426.html(101,966 bytes); the three PDFs, 19,168,256, 19,111,775 and 19,410,544 bytes, with their extracted text beside them; the archived department pages named above.
Sources
AusTender, contract notice CN4118426, Department of Employment and Workplace Relations, "Assurance Review to support the Targeted Compliance Framework", https://www.tenders.gov.au/Cn/Show/05c19be8-a95f-4379-bd35-630cc2f472c6
Department of Employment and Workplace Relations, "Targeted Compliance Framework Assurance Review – Final Report" (page and PDF, three versions), https://www.dewr.gov.au/assuring-integrity-targeted-compliance-framework/resources/targeted-compliance-framework-assurance-review-final-report; Internet Archive captures of the page (22 August 2026) and of the download address (16 August 2025, 6 October 2025, 28 July 2026)
Department of Employment and Workplace Relations, Secretary's "Statement on progress under the Targeted Compliance Framework Integrity Assurance Program", 3 October 2025, https://www.dewr.gov.au/assuring-integrity-targeted-compliance-framework/announcements/statement-secretary-progress-under-targeted-compliance-framework-integrity-assurance-program (Internet Archive capture, 3 October 2025)
Department of Employment and Workplace Relations, "Targeted Compliance Framework Assurance Review – Statement of Assurance" (page note: "This Statement was updated on 26 September 2025 and replaces the Statement of Assurance dated 18 June 2025"), https://www.dewr.gov.au/assuring-integrity-targeted-compliance-framework/resources/targeted-compliance-framework-assurance-review-statement-assurance
Rod McGuirk, "Deloitte to partially refund Australian government for report with apparent AI-generated errors", Associated Press, 7 October 2025
Aaron Patrick, "Law lecturer Christopher Rudge slams Deloitte's government-funded report written with AI", The Nightly, 6 October 2025
Australian Financial Review, "Academics raise alarm over suspected AI use in Deloitte report", 22 August 2025, and "Deloitte report suspected of containing AI invented quote from Robo-debt case", 25 August 2025 (dates as indexed by the AI Incident Database, incident 1193; the articles are paywalled)
Chris Johnson, "Government urged to get Deloitte's total fee returned for delivering a botched AI report", PS News, 12 October 2025
The Australian Greens, "Greens slam Deloitte's unethical behaviour and measly refund over AI report bungle", media release, 9 October 2025
Alexei Alexis, "Deloitte refunds over $60K for report with AI errors, Australian government says", CFO Dive, 21 October 2025
Simon Willison, "Deloitte to pay money back to Albanese government after using AI in $440,000 report", 6 October 2025 (quoting The Guardian of the same date and the report's page 2 note)
