In April 2026 South Africa withdrew its draft national AI policy after it was found to cite sources that do not exist. In the year before, invented citations surfaced in a White House report, in consulting reports commissioned by governments in Australia and Canada, in research published by large professional-services firms, in expert evidence, and invented quotations and misquoted law in orders drafted in two US federal judges’ chambers. In at least nine of the eighteen cases the record describes a review the document went through: partners, cite-checkers, editors, peer reviewers or a Cabinet; in one, covering two judges’ orders, the review that existed was bypassed.
A document can pass several reviews without anyone establishing that its references exist and support its claims.
What this covers: reports, filings, research and policy documents whose references were invented or misattributed, most of them with generative AI involved by the author’s own account, the finder’s evidence or a court’s finding. It matters to anyone who writes, commissions, publishes or relies on such work. See What We Mean by AI.
The cases
Eighteen documented cases, 2025 to 2026, newest first. Open a case to see what was wrong, the review it passed and what happened next.
Who found the problem: A party to the case, its counsel or another firm 7 · Journalists 3 · Academics 2 · AI-detection vendor (GPTZero) 3 · Member of the public 1 · Court-appointed special master 1 · Not stated 1
LeDoux v. Outliers (US federal court, Washington) · July and August 2026 · Two expert reports and several court filings
Found by: The opposing party
- What was wrong
- Dozens of inaccurate citations, including citations to sources that do not exist and invented quotations; two expert reports cited the same nonexistent articles.
- What review was in place
- Counsel said she used Claude and ChatGPT and did not check their output against the sources. Both experts signed reports built on the same AI-generated citation table and neither caught the errors.
- What happened next
- Counsel was sanctioned US$3,000; an expert was excluded and the client’s remaining claims were dismissed.
PwC Middle East · July 2026 · Four published research reports, 2024 to 2026
Found by: GPTZero, an AI-detection company, reported in the press
- What was wrong
- According to GPTZero’s analysis, reported by the International Accounting Bulletin, fabricated or unsupported citations, including an academic paper on air quality in Riyadh that appeared to have been invented.
- What review was in place
- Not described by the firm.
- What happened next
- PwC Middle East said it was updating a limited number of supporting citations.
KPMG · June 2026 · A published report on agentic AI in customer experience
Found by: GPTZero, then the Financial Times
- What was wrong
- GPTZero found that only 5 of 45 citations pointed accurately to real sources; organizations named in the report told the Financial Times its claims about them were untrue or misleading.
- What review was in place
- KPMG said its guidelines require human oversight to validate content and verify sources.
- What happened next
- KPMG removed the report while it investigated.
EY Canada · May 2026 · A published report on fraud in loyalty programs
Found by: GPTZero, through Sherwood News
- What was wrong
- According to GPTZero, most of its 27 references were AI-generated and made up or referenced incorrectly.
- What review was in place
- Credited to two partners and a senior manager; EY said it was examining how the report was approved.
- What happened next
- EY removed the report.
South Africa, Department of Home Affairs · April 2026 · A Cabinet-approved white paper on citizenship and immigration
Found by: Not stated by the department
- What was wrong
- Apparent AI hallucinations among the references appended to the paper.
- What review was in place
- Approved by Cabinet; the department said the references were attached after the fact and are not cited in the body.
- What happened next
- Two senior officials suspended, the reference list withdrawn, and law firms appointed to review every policy document since November 2022; AI checks and declarations are to be added to approvals.
South Africa, Draft National AI Policy · April 2026 · A draft national policy published for comment
Found by: News24
- What was wrong
- Fictitious sources in the reference list.
- What review was in place
- The minister said AI-generated citations were most plausibly included without proper verification, and that his department had not met the standard expected of it.
- What happened next
- The draft was withdrawn on April 26, 2026.
Sullivan & Cromwell · April 2026 · An emergency motion in a US bankruptcy court
Found by: Another law firm
- What was wrong
- AI-generated fabricated or inaccurate citations.
- What review was in place
- Citation-review policies, two tracked training modules before any AI access, and a rule to check all AI output independently; the firm said its review process did not identify the errors.
- What happened next
- The signing partner thanked the other firm for raising the errors; a re-review of the firm’s other filings found no further AI errors.
University of Hong Kong · December 2025 · A peer-reviewed journal article
Found by: A member of the public, on Threads
- What was wrong
- At least 24 references the publisher could not verify; AI use was not disclosed.
- What review was in place
- The corresponding author said the paper passed two rounds of review; the journal acknowledged an oversight in checking references.
- What happened next
- The paper was retracted and the corresponding author resigned as associate dean.
Deloitte Canada · November 2025 · A health workforce plan for Newfoundland and Labrador
Found by: The Independent
- What was wrong
- At least four citations that do not, or appear not to, exist.
- What review was in place
- Deloitte said AI was used selectively to support a small number of research citations, not to write the report.
- What happened next
- Deloitte acknowledged four incorrect citations, stood by its findings and committed to a full citation review.
Education Accord NL · September 2025 · A 10-year education plan for Newfoundland and Labrador
Found by: Memorial University academics, via Radio-Canada
- What was wrong
- At least 15 citations to journal articles and documents that do not exist.
- What review was in place
- A co-chair said “We checked everything”; another placed the failure in editing and reference checking.
- What happened next
- Pulled from the government website; a corrected version had not been posted by September 2026.
Deloitte Australia · August 2025 · An assurance review for a federal department
Found by: A University of Sydney academic, through the media
- What was wrong
- Nonexistent references and citations.
- What review was in place
- Deloitte told the government its product was reviewed before delivery and that the review corrected some errors but not all.
- What happened next
- Deloitte repaid the final instalment of A$97,587.11 and the report was corrected.
Two US federal judges · June and July 2025 · Court orders
Found by: The parties in the cases
- What was wrong
- Orders that misquoted law, named people who were not in the case and attributed invented quotations to parties.
- What review was in place
- Both chambers said drafts normally pass several levels of review, including cite-checking tools; the orders were early drafts docketed without it.
- What happened next
- Both orders were removed from the docket and one was replaced three days later; one judge now requires a second clerk to review every draft.
Johnson v. Dunn (Butler Snow) · July 2025 · Court motions
Found by: The opposing party
- What was wrong
- Citations fabricated by ChatGPT.
- What review was in place
- A partner added the citations without checking them; the lawyer who signed checked only grammar; the practice-group leader gave one motion only a brief scan.
- What happened next
- Three lawyers publicly reprimanded, disqualified and referred to the state bar.
White House MAHA report · May 2025 · A federal government report
Found by: NOTUS
- What was wrong
- Seven of 522 citations were to papers that could not be found.
- What review was in place
- None described.
- What happened next
- Citations were corrected; the White House called them “formatting issues”.
Concord Music v. Anthropic (Latham & Watkins) · May 2025 · An expert declaration
Found by: The opposing party
- What was wrong
- A citation formatted by Claude with the wrong title and authors.
- What review was in place
- The expert and counsel reviewed each reference, and a manual cite-check confirmed the source and corrected page numbers but missed the wrong title and authors.
- What happened next
- The court struck the paragraph that relied on the citation and said the issue undermined the declaration’s credibility; the firm added multiple levels of review.
Lacey v. State Farm · May 2025 · A brief by two law firms
Found by: The court-appointed special master
- What was wrong
- About 9 of 27 citations were wrong in some way; at least two authorities did not exist.
- What review was in place
- No one at either firm cite-checked an AI-produced outline passed between them; a corrected resubmission kept most of the errors.
- What happened next
- The briefs were struck and the firms ordered to pay US$31,100.
Reddy v Saroya (Alberta Court of Appeal) · 2025 and January 2026 · An appeal factum drafted by a contractor
Found by: The opposing party’s counsel
- What was wrong
- Seven cited cases could not be found.
- What review was in place
- Drafted by a third-party contractor; the court held the lawyer named on the filing responsible.
- What happened next
- In January 2026 the court ordered costs of C$17,550 plus GST against lead counsel personally.
Kohls v. Ellison · January 2025 · An expert declaration for a state attorney general
Found by: The opposing party
- What was wrong
- Two nonexistent academic articles and one misattribution, from GPT-4o.
- What review was in place
- The attorney general’s office said it had no idea the declaration contained AI-generated citations.
- What happened next
- The declaration was excluded.
Three of the cases rest on investigations by GPTZero, which sells AI-detection software and so has a commercial interest in finding such errors; in each, the firm removed the report or said it was correcting citations. A firm’s response is its own account; KPMG and Deloitte had not published the outcome of their reviews by October 2026.
The reviews were there
The striking feature of these cases is not that no one checked. It is what the checks were looking for.
- Deloitte told the Australian government that its product was reviewed before delivery, and that the review corrected some errors but did not identify all of them.
- Sullivan & Cromwell said its lawyers must complete two tracked training modules before using AI, that its rules require independent checks of AI output, and that its citation review still did not identify the AI-generated errors.
- Latham & Watkins’ manual cite-check confirmed that a source supported the point and corrected its page numbers, but missed that the title and authors were wrong; the court said it was not clear how a complete change of title could escape a human cite-check.
- Two federal judges said drafts in their chambers normally pass several levels of review, including cite-checking tools, but that the orders in question were early drafts docketed without it.
- A University of Hong Kong paper passed two rounds of review, according to its corresponding author, and the journal acknowledged an oversight in checking its references.
- Education Accord NL’s co-chairs said they had submitted a checked report and that the failure was in editing and reference checking; one said “We checked everything”.
Review built for human error checks that a reference is formatted correctly and supports the point. AI can produce references that look correct and do not exist, or that point to a real document under the wrong title and authors, so the check that works is the one that opens the source and compares every element of the citation with it. In the Australian case, the department told a Senate committee that quality assurance for AI-assisted work needs to be framed to pick up what AI might get wrong.
Who found them
In every case where the record says who found the problem, it was someone other than the people who produced the document. A party to the case, its counsel or another firm found seven; checking the other side’s citations is part of their work. Journalists, academics, a detection company and a member of the public found most of the rest. In the Deloitte Australia case, the department learned of the errors through a media request.
What it cost
- Money. Deloitte repaid the final instalment of A$97,587.11 on its contract with the Australian department, in a repayment confirmed in October 2025. In May 2025 two law firms were ordered to pay US$31,100 to the other side, and in July 2026 a lawyer was sanctioned US$3,000 personally.
- The work itself. An expert’s evidence was excluded and his client’s case dismissed; another expert declaration was excluded; a paper was retracted; a national policy draft was withdrawn; reports were removed from firms’ websites.
- People. South Africa’s Home Affairs department suspended two senior officials and ordered a review of every policy document it had produced since November 2022. Three lawyers were reprimanded, disqualified and referred to their state bar.
What would have prevented it
| What happened | What would have caught it | Evidence from the cases |
|---|---|---|
| The source was opened and confirmed to support the point, but the citation’s title and authors were not compared with it | Open every cited source and confirm it exists, that its title, authors and details match the citation, and that it says what is claimed, before release | A cite-check that confirmed format missed a wrong title and authors |
| Senior lawyers’ names were on the filings; one did not review one motion and only scanned the other, and the lawyer who signed checked only grammar | Make the person who signs responsible for having checked, and record what they checked | The lawyer named on a filing bears responsibility even when a contractor drafted it |
| AI output passed between people and firms unchecked | Tell the next reviewer when AI was used, and check at every handover | The receiving firm did not know AI had been used and did not ask |
| Drafts went out without the review that existed | Make the check a gate in the release process, not a habit | One judge now requires a second clerk’s review of every draft |
| Training and policy were in place and not followed | Train people on what AI errors look like, and test whether the check happened | A department now plans AI checks and declarations as part of approval |
Are you at risk?
For anyone who writes, commissions or relies on reports, filings or research.
- Does anyone open each cited source before your work is released, or does review check only that references look complete?
- Do the people who sign the work know which parts were drafted with AI?
- When AI-assisted work passes between colleagues, firms or contractors, does each handover include a check?
- Is the citation check a required step before release, with a record that it was done?
- When you commission work, does the contract require disclosure of AI use and say who checks the references?
- If an outsider found an invented reference in your work tomorrow, could you show what was checked?
Sources and further reading
- SAnews (South African Government News Agency), “Minister announces withdrawal of draft AI Policy”, April 26, 2026.https://www.sanews.gov.za/south-africa/minister-announces-withdrawal-draft-ai-policy
- Department of Home Affairs, South Africa, “Home Affairs suspends two officials over AI use linked to revised White Paper”, April 30, 2026.https://www.gov.za/news/media-statements/home-affairs-suspends-two-officials-over-ai-use-linked-revised-white-paper
- Parliament of Australia, Education and Employment Legislation Committee, Estimates, Hansard, October 9, 2025.https://parlinfo.aph.gov.au/parlInfo/search/display/display.w3p;query=Id%3A%22committees%2Festimate%2F28997%2F0002%22
- U.S. District Court for the Western District of Washington, LeDoux v. Outliers, Inc., order on sanctions, July 24, 2026.https://storage.courtlistener.com/recap/gov.uscourts.wawd.339711/gov.uscourts.wawd.339711.265.0.pdf
- U.S. District Court for the Western District of Washington, LeDoux v. Outliers, Inc., order on summary judgment, August 18, 2026.https://storage.courtlistener.com/recap/gov.uscourts.wawd.339711/gov.uscourts.wawd.339711.269.0.pdf
- Sullivan & Cromwell, letter to Chief Judge Martin Glenn, U.S. Bankruptcy Court for the Southern District of New York, April 18, 2026.https://storage.courtlistener.com/recap/gov.uscourts.nysb.334560/gov.uscourts.nysb.334560.25.0.pdf
- U.S. District Court for the Northern District of California, Concord Music Group, Inc. v. Anthropic PBC, order, May 23, 2025.https://storage.courtlistener.com/recap/gov.uscourts.cand.431519/gov.uscourts.cand.431519.377.0.pdf
- U.S. District Court for the Central District of California, Lacey v. State Farm General Insurance Co., special master’s order, May 5, 2025.https://storage.courtlistener.com/recap/gov.uscourts.cacd.930490/gov.uscourts.cacd.930490.119.0.pdf
- U.S. Senate Committee on the Judiciary, “Grassley Releases Judges’ Responses Owning Up to AI Use”, October 23, 2025.https://www.judiciary.senate.gov/press/rep/releases/grassley-releases-judges-responses-owning-up-to-ai-use-calls-for-continued-oversight-and-regulation
- CBC News, “N.L.’s 10-year education action plan cites sources that don’t exist”, September 12, 2025.https://www.cbc.ca/news/canada/newfoundland-labrador/education-accord-nl-sources-dont-exist-1.7631364
- Hong Kong Free Press, “University of Hong Kong professor steps down from associate deanship after AI-generated references scandal”, December 18, 2025.https://hongkongfp.com/2025/12/18/university-of-hong-kong-professor-steps-down-from-associate-deanship-after-ai-generated-references-scandal/
- NBC News, “Trump admin corrects RFK Jr.’s MAHA report after citation errors”, May 30, 2025.https://www.nbcnews.com/health/health-news/trump-admin-corrects-rfk-jrs-maha-report-citation-errors-rcna209913
- GPTZero, “Chasing the Hallucinations: KPMG’s AI-Powered Attempt at ‘Redefining Excellence’”, June 12, 2026.https://gptzero.me/news/investigations-kpmg/
- U.S. District Court for the Northern District of Alabama, Johnson v. Dunn, No. 2:21-cv-1701, sanctions order, July 23, 2025.https://storage.courtlistener.com/recap/gov.uscourts.alnd.179677/gov.uscourts.alnd.179677.204.0.pdf
- U.S. District Court for the District of Minnesota, Kohls v. Ellison, No. 24-cv-3754, order on motion to exclude, January 10, 2025.https://storage.courtlistener.com/recap/gov.uscourts.mnd.220348/gov.uscourts.mnd.220348.46.0.pdf
- U.S. District Court for the Northern District of California, Concord Music Group, Inc. v. Anthropic PBC, No. 5:24-cv-03811, declaration of counsel, May 15, 2025.https://storage.courtlistener.com/recap/gov.uscourts.cand.431519/gov.uscourts.cand.431519.371.0.pdf
- Judge Julien X. Neals, letter to the Director of the Administrative Office of the U.S. Courts, October 20, 2025.https://www.grassley.senate.gov/imo/media/doc/neals_to_grassley_re_judiciary_use_of_ai.pdf
- Judge Henry T. Wingate, letter to the Director of the Administrative Office of the U.S. Courts, October 21, 2025.https://www.grassley.senate.gov/imo/media/doc/wingate_to_grassley_re_judiciary_use_of_ai.pdf
- ABlawg, “The Alberta Court of Appeal Weighs in on the use of AI in Court Submissions”, October 15, 2025.https://ablawg.ca/2025/10/15/the-alberta-court-of-appeal-weighs-in-on-the-use-of-ai-in-court-submissions/
- CBC News, “Education Accord co-chairs blame government for fake sources”, September 17, 2025.https://www.cbc.ca/news/canada/newfoundland-labrador/nl-education-accord-goodnough-burke-1.7635123
- TimesLIVE, “Minister Malatsi withdraws draft AI policy”, April 27, 2026.https://www.timeslive.co.za/news/2026-04-27-minister-malatsi-withdraws-draft-ai-policy/
- The Guardian, “Deloitte to pay money back to Albanese government after using AI in $440,000 report”, October 6, 2025.https://www.theguardian.com/australia-news/2025/oct/06/deloitte-to-pay-money-back-to-albanese-government-after-using-ai-in-440000-report
- The Associated Press, “Deloitte to partially refund Australian government for report with apparent AI-generated errors”, October 7, 2025.https://www.yahoo.com/news/articles/deloitte-partially-refund-australian-government-070855665.html
- The Independent, “Major N.L. healthcare report contains errors likely generated by A.I.”, November 22, 2025.https://theindependent.ca/news/lji/major-n-l-healthcare-report-contains-errors-likely-generated-by-a-i/
- The Independent, “Deloitte breaks silence on N.L. healthcare report”, November 26, 2025.https://theindependent.ca/news/lji/deloitte-breaks-silence-on-n-l-healthcare-report/
- Information Age, “EY retracts cyber report littered with AI errors”, May 19, 2026.https://ia.acs.org.au/article/2026/ey-retracts-cyber-report-littered-with-ai-errors.html
- International Accounting Bulletin, “EY removes loyalty rewards study after AI hallucinations found”, May 18, 2026.https://www.internationalaccountingbulletin.com/news/ey-removes-loyalty-rewards-study-after-ai-hallucinations-found/
- Sherwood News, “AI hallucinations appear to be creeping into consulting reports”, May 14, 2026.https://sherwood.news/tech/ai-hallucinations-appear-to-be-creeping-into-consulting-reports/
- TechCrunch, “KPMG pulls report on AI usage due to apparent hallucinations”, June 13, 2026.https://techcrunch.com/2026/06/13/kpmg-pulls-report-on-ai-usage-due-to-apparent-hallucinations/
- GPTZero, “Chasing the Hallucinations: PwC report hallucinates product and government customers”, July 29, 2026.https://gptzero.me/news/investigations-pwc/
- International Accounting Bulletin, “PwC AI reports tainted by hallucination errors – report”, July 31, 2026.https://www.internationalaccountingbulletin.com/news/pwc-ai-reports-tainted-by-hallucination-errors-report/
- City AM, “PwC thought leadership reports ‘100 per cent AI generated’”, July 29, 2026.https://www.cityam.com/pwc-thought-leadership-reports-found-to-contain-ai-hallucinations/
- GPTZero, “Investigation: Hallucinations in Ernst & Young Report on Loyalty Fraud”, May 14, 2026.https://gptzero.me/investigations/ey
- GPTZero, “AI Detector”, company home page.https://gptzero.me/
- Court of Appeal of Alberta, Reddy v Saroya, 2026 ABCA 20, memorandum of judgment regarding costs, January 22, 2026.https://www.canadianlawyermag.com/cl-plus/cases/download/13133/9656/Reddy%20v%20Saroya,%202026%20ABCA%2020.pdf
- CBC News, “‘Worse than we thought’: Scrubbing of Education Accord reveals 38 references to non-existent sources”, September 3, 2026.https://www.cbc.ca/news/canada/newfoundland-labrador/education-accord-report-scrubbed-9.7329683
- CBC News, “N.L. asks Deloitte to carry out review after ‘incorrect’ citations found in $1.6M provincial health plan”, November 24, 2025.https://www.cbc.ca/news/canada/newfoundland-labrador/nl-deloitte-citations-9.6990216
Related reading
Review status
Orientation, not legal advice. Each statement about a named organization rests on a court record, a government statement, the organization’s own statement or a named publication, and goes no further than that source; the investigations by GPTZero are identified as a detection vendor’s own analysis.