UKWELI.Integrity Watch
← Home Updated weekly · 5 Aug 2026 Integrify your document
The State of Document Integrity

The documents are failing. We keep the receipts.

A verified, continuously updated log of real incidents where reports, filings and policies were withdrawn, refunded, corrected or sanctioned because they contained AI fabrications. Every entry is sourced. Nothing here is an opinion.

Every entry verified against a reputable source · logged weekly, never padded.
The register
72
verified incidents across government, the Big Four, the courts and the academic record — and counting.
Latest: High Court of South Africa, Gauteng Division, Johannesburg (Acting Judge Themba Khaba) · 5 days ago
1 in 277
of biomedical papers in early 2026 cites a fabricated reference — a 12× rise in two years.
2023
1 in 2,828
2025
1 in 458
EARLY 2026
1 in 277
The sharp rise began mid-2024, as AI writing tools spread.
Source: The Lancet · Columbia Nursing audit of 97.1M references · May 2026
The scale

It is not happening to amateurs.

Each figure below is a single, sourced fact — not a total. We do not add refunds, fines and contract values together; that would be exactly the kind of unearned number Ukweli exists to catch.

KPMG · Deloitte · EY
Three of the Big Four published reports with AI fabrications — alongside the UK High Court and a national government.
See the register below for each, with sources.
$110,000
The largest single court penalty in the log for one set of fabricated citations — a US federal sanction.
ABA Journal · District of Oregon, May 2026
97.1M references
Audited in the biomedical literature; 1 in 277 papers carries a fabricated reference, up 12× in two years.
The Lancet · Columbia Nursing, May 2026
The register

Every case. Verified. Sourced.

The trend so far · 72 verified incidents by year
010203040Legal: 33% of 2023 (1 of 3)Government: 33% of 2023 (1 of 3)Media: 33% of 2023 (1 of 3)32023Legal: 71% of 2024 (5 of 7)Government: 14% of 2024 (1 of 7)Media: 14% of 2024 (1 of 7)72024Legal: 52% of 2025 (16 of 31)Government: 13% of 2025 (4 of 31)Consulting: 10% of 2025 (3 of 31)Academic: 13% of 2025 (4 of 31)Media: 13% of 2025 (4 of 31)312025Legal: 61% of 2026 (19 of 31)Government: 10% of 2026 (3 of 31)Consulting: 3% of 2026 (1 of 31)Academic: 10% of 2026 (3 of 31)Media: 16% of 2026 (5 of 31)312026Jan–Aug
LegalGovernmentConsultingAcademicMedia

Verified incidents in the register, stacked by sector. 2026 is year-to-date (dashed). This counts the register itself — densest for recent, English-language cases — not the true global incidence rate (that is the Lancet figure above).

By type & consequence — click a bar to filter
By fabrication type
By consequence
Severity runs low to high: Corrected → Refunded → Withdrawn → Sanctioned. One fabricated citation has been enough to trigger each.

Filter by sector, region, consequence or year — or search. Newest first. Each entry links to its source and has its own permalink.

Is your sector exposed?
Pick your sector to see the verified incidents in your world.
Showing 72 of 72 verified incidents

High Court of South Africa, Gauteng Division, Johannesburg (Acting Judge Themba Khaba)

Legal
South Africa
Legal authorities +1
12 tabulated citation discrepancies in a High Court judgment, including a case that does not exist
On appeal, the full bench tabulated 12 numbered citation discrepancies in the Acting Judge's judgment (the judgment text and the reporting both refer to "11 discrepancies"). The standout is "Lubbe v Volkswagon SA [2023] 4 All SA 289 (C)" — "This case does not exist" — relied on in five paragraphs: [75], [92], [93], [108] and [112]. Others include a citation belonging to an unrelated case (Myeni given SA Express Ltd v Bagport's citation; MEC for Health Eastern Cape v Mkhintha given Notshokovu v S's), quotes attributed to a correctly named case that in fact come from a different, uncited judgment, and a citation that does not exist at all (Premier for the Province of Gauteng [2020] 4 All SA 559 (SCA)). The discrepancies were identified by the appellant's legal team, agreed by the respondent's, independently confirmed by the senior librarian of the Johannesburg Society of Advocates, and separately verified by a court researcher before the hearing. Critically, the judgment recorded in some instances that junior counsel Ms K Mitchell had supplied references she had never provided: "The errors can therefore not be attributed to counsel."
Sanctioned 31 Jul 2026 · source ↗

US District Court, Western District of Washington (Judge Tiffany M. Cartwright)

Legal
United States
Legal authorities +2
Dozens of false citations across 4 briefs and 2 expert reports
Dozens of inaccurate factual and legal citations across at least five filings — four briefs and two expert reports — including citations to nonexistent sources and fictitious quotations attributed to real, existing sources. Counsel generated a formatted citation table using Claude and ChatGPT and supplied the same AI-generated table to both testifying experts as an appendix; neither expert caught the errors before signing.
Sanctioned 24 Jul 2026 · source ↗

US Bankruptcy Court, Southern District of Texas

Legal
United States
Legal authorities +1
Sanctioned over fabrications produced with a paid legal-research AI
PER THE COURT'S ORDER AS PUBLISHED BY THE AI HALLUCINATION CASES DATABASE, AND NOT CORROBORATED BY ANY INDEPENDENT REPORT AS AT 5 AUGUST 2026: motions to quash filed by counsel for the petitioning creditors contained multiple fabricated quotations, citations to non-existent authority, and misleading descriptions of real authority. Counsel — a sole practitioner — said he had not used ChatGPT but had used the generative-AI feature in Westlaw's 'Precision' product; the court rejected his argument that a legal AI tool should be treated more leniently than a consumer one.
Sanctioned 14 Jul 2026 · source ↗

UK Home Office

Government
United Kingdom
Sources & experts +1
An asylum refusal built on a policy document that does not exist
The Home Office refused a Moroccan woman's asylum claim citing a country policy and information note (CPIN) on Morocco dated July 2021 as authority that Morocco would be safe for her return. The Upper Tribunal could not locate the document and the Home Office's own country policy and information team was unable to confirm it existed; the judge held "it appears that no such CPIN has ever existed." No CPIN on Morocco has been published since 2017. The First-tier Tribunal had already dismissed her appeal relying on the same non-existent document. Asked to produce it, the Home Office first sent a link to an archive where it could not be found, then produced a different, unpublished internal guidance document (a country information note, or CIN) that had never been disclosed to the court — and which does not conclude Morocco is safe for the forcible return of women in her position, recording instead that 40,000 child marriages were granted by exemption in 2018 (one-fifth of all marriages that year) and that Moroccan law does not explicitly criminalise marital rape.
Withdrawn 14 Jul 2026 · source ↗

Tribunal Superior de Xustiza de Galicia (Sala do Social, A Coruña)

Legal
Spain
Legal authorities +1
24 of 32 citations (75%) fabricated
Of 32 case citations in the appeal, 24 (75%) were fabricated or wrong — non-existent Supreme Court, Constitutional Court and Court of Justice of the EU rulings, several with invented verbatim quotations, all verified false against the CENDOJ case database. The lawyer used a free generative-AI tool without verification and admitted responsibility. Independently corroborated 5 Aug 2026 by four Spanish outlets, all reporting 24 fabricated references and the EUR1,800 fine; Iberley additionally confirms the auto reference (Auto n. 00061/2026, A Coruna, 7 July 2026, recurso de suplicacion 5472/2025) and the 75% proportion. The raw total of 32 is stated in the auto and follows arithmetically from the two corroborated figures.
Sanctioned 7 Jul 2026 · source ↗

Supreme Court of India

Legal
India
Legal authorities +1
3 of 6 cited precedents did not exist
Of six precedents relied on by the NCLT, three did not exist and three others carried invented paragraphs or an incorrect case title; an affidavit confirmed the authorities could not be traced in any recognised legal database.
Withdrawn 2 Jul 2026 · source ↗

Labour Court of South Africa, Durban; fabrications originated with the chairperson of a Matjhabeng Local Municipality internal disciplinary hearing

Legal
South Africa
Legal authorities
An adjudicator's own ruling cited authorities that do not exist
The chairperson of an internal disciplinary hearing issued a ruling citing case authorities that do not exist. This is established from the primary record: the Gauteng High Court, in FJL v TGO (2025/220239, 31 July 2026, reportable), cites the case at footnote 9 as (2026/157576) [2026] ZALCD 30 and records at paragraph [40] that a stay was granted where "rulings made by a disciplinary chairperson (not a Judge or Acting Judge) citing case authorities which did not exist" were under challenge. PER CLIFFE DEKKER HOFMEYR'S EMPLOYMENT LAW ALERT OF 27 JULY 2026, AND NOT YET CHECKED AGAINST THE JUDGMENT ITSELF: the ruling relied on four authorities, of which only one existed and supported the proposition cited, two did not exist at all, and a fourth was a real decision that did not address the interpretation of collective agreements as claimed. Those figures are single-sourced to that alert and are attributed rather than asserted.
Withdrawn Jul 2026 · source ↗

KPMG

Consulting
United Kingdom
References & studies +1
40 of 45 citations fabricated
Of 45 citations, only 5 pointed to real, intact sources; 40 of 45 titles were fabricated. Case studies about UBS, the UK NHS, Swiss Federal Railways and Transport for London were disputed as inaccurate by the named organisations.
Withdrawn 13 Jun 2026 · source ↗

US District Court, Northern District of Mississippi (Judge Sharion Aycock)

Legal
United States
Legal authorities
Both sides hallucinated — all 4 attorneys removed
Hallucinated case citations across three briefs filed by both opposing sides in the same matter; the court, not opposing counsel, discovered them. One out-of-state attorney admitted she had filed further pleadings before the same court containing undiscovered fictitious citations, and had continued citing hallucinated cases in other jurisdictions after the problem surfaced. The order does not state a total citation count.
Sanctioned 8 Jun 2026 · source ↗

Oregon Supreme Court

Legal
United States
Legal authorities +1
First state-Supreme-Court action
Citations to prior cases that do not exist in the Oregon Appellate Reports or the Oregon Reports, plus fabricated quotes; a later declaration cited at least four more non-existent cases.
Sanctioned 5 Jun 2026 · source ↗

US Court of Appeals for the Ninth Circuit

Legal
United States
Legal authorities
Two attorneys suspended 6 months
Fabricated citations; the attorneys characterised the errors as typographical mistakes that did not involve AI.
Sanctioned 4 Jun 2026 · source ↗

New York Supreme Court, Appellate Division, Second Department

Legal
United States
Legal authorities +1
$10,500 sanction
The brief cited several non-existent cases, included fictitious purported Court of Appeals quotations “wholly contrary to actual law,” and misrepresented the holdings of real cases.
Sanctioned Jun 2026 · source ↗

Journal of Medical Ethics (BMJ Group)

Academic
United Kingdom
References & studies
References that do not exist, plus peer-review manipulation
The journal's investigation found several cited references did not exist. The sole author — confirmed to be a high-school student (affiliation listed as Shrewsbury Public Schools, Massachusetts) — used generative AI to "identify and understand referenced sources" and did not verify the references before submission. The retraction notice also cites "evidence of peer review manipulation."
Withdrawn 28 May 2026 · source ↗

Berlingske

Media
Denmark
Quotes +1
Invented professor and fabricated quotes in a national daily
An internal AI tool used to summarise expert quotes was run without the source material and invented both quotations and people — including a non-existent DTU technology professor, 'Anne Kjær Nielsen.' Fabricated quotes were attributed to sources who do not exist. Flagged by a reporter at Dagbladet Information.
Corrected 27 May 2026 · source ↗

Steven Rosenbaum (author)

Media
United States
Quotes +1
Half-a-dozen-plus fabricated quotes in a book about AI and truth
A New York Times review found more than half a dozen misattributed or fabricated quotes. One attributed to tech journalist Kara Swisher ("the most sophisticated AI language model is like a mirror...") was disowned by her, who said she "never said that"; further misattributed quotes involved other named figures. The author disclosed using ChatGPT and Claude during research, writing and editing.
Corrected 19 May 2026 · source ↗

US District Court, Eastern District of North Carolina (Magistrate Judge Robert T. Numbers II) / US Department of Justice

Legal
United States
Quotes +1
US federal prosecutor fired over an AI-fabricated filing
A brief the federal prosecutor drafted with a generative-AI tool and filed without verifying contained fabricated quotations, misstatements of case holdings, and fabricated regulatory language. The pro se plaintiff — a retired military lawyer — noticed the brief did not read like the sources it cited.
Sanctioned 28 Apr 2026 · source ↗

DCDT (SA Department of Communications & Digital Technologies)

Government
South Africa
References & studies
Fake academic citations in a national policy
At least 6 of 67 academic citations were AI-generated hallucinations (fake journals / articles never published).
Withdrawn 26 Apr 2026 · source ↗

Alabama Supreme Court

Legal
United States
Legal authorities
Appeal dismissed; $17,200 in costs
Numerous invalid, irrelevant and outright fabricated citations, many AI-generated and to cases that do not exist; the apologising footnote cited two further non-existent cases.
Sanctioned 24 Apr 2026 · source ↗

Nebraska Supreme Court

Legal
United States
Legal authorities +1
First US attorney suspended from practice over AI filings
A divorce-appeal brief the court found "potentially AI-generated" was "filled with fictitious cases … and fictitious quotations from statutes." The court listed 20 examples of fictional quotes, incorrect case numbers, or misrepresented holdings, including a wholly fictitious 2019 case, Kennedy v. Kennedy, cited with fabricated quotations. By opposing counsel's count, 57 of 63 references in the brief were defective.
Sanctioned 16 Apr 2026 · source ↗

The New York Times

Media
United States
Quotes
An AI summary published as a direct quote
A direct quotation attributed to Conservative leader Pierre Poilievre was, per the Times' own editor's note, an AI-generated summary of his views that the AI tool rendered as a verbatim quotation. Poilievre never said the quoted words; the note stated the reporter "should have checked the accuracy of what the A.I. tool returned."
Corrected 14 Apr 2026 · source ↗

Mediahuis (journalist Peter Vandermeersch)

Media
Netherlands
Quotes +1
Fabricated quotes in 15 of 53 posts
In 15 of 53 blog posts, quotes were attributed to named individuals and to publications where they could not be found; seven named individuals confirmed they never made the statements — dozens of fabricated quotes in total.
Withdrawn 19 Mar 2026 · source ↗

Tribunal Superior do Trabalho (TST — Brazil)

Legal
Brazil
Legal authorities +1
Fake precedents attributed to real ministers
Cited non-existent precedents, including ones falsely attributed to a sitting and a retired Minister (one dated after his retirement); none existed in the TST system.
Sanctioned 10 Mar 2026 · source ↗

Digestive Diseases and Sciences (Springer Nature)

Academic
United Kingdom
References & studies
12 of 14 references do not exist
12 of 14 references found not to exist; recent, easily-locatable citations could not be traced.
Withdrawn 6 Mar 2026 · source ↗

High Court of Kenya, Milimani Law Courts (Justice Bahati Mwamuye)

Legal
Kenya
Legal authorities
AI-drafted court filing struck out
The Notice of Motion and supporting affidavit were entirely machine/AI-generated and contained citations to non-existent cases.
Withdrawn 6 Mar 2026 · source ↗

US Court of Appeals for the Sixth Circuit

Legal
United States
Legal authorities
$30K, 24+ bad cites
More than two dozen citations across three consolidated appeals were incorrect, misrepresented, or entirely non-existent.
Sanctioned Mar 2026 · source ↗

West Midlands Police

Government
United Kingdom
Claims & data +1
One fictitious fixture, used to classify away fans as high-risk
Microsoft Copilot generated a fixture between Maccabi Tel Aviv and West Ham that never took place; the non-existent match was used in the force's intelligence report as evidence that Maccabi fans posed an unusual risk. A 10 October 2025 reply from the operation's Gold Commander carried further AI-sourced claims about the Amsterdam disorder — that 2,000 Dutch police had been deployed, that “people were thrown into the river”, and that the disorder was “well organised and targeted towards Muslim communities”.
Corrected 22 Feb 2026 · source ↗

Corte Suprema de Justicia de Colombia

Legal
Colombia
Legal authorities
~US$6,000 fine for fabricated authorities
A legal appeal contained non-existent legal provisions and precedents generated by an AI tool.
Sanctioned 16 Feb 2026 · source ↗

Ars Technica

Media
United States
Quotes
AI-faked quote; reporter out
A direct quotation attributed to a real, named source he never said — an AI paraphrase used as a verbatim quote.
Withdrawn 13 Feb 2026 · source ↗

Intensive Care Medicine (Springer)

Academic
Netherlands
References & studies
Fake reference to its own journal
Among 15 references, a fabricated citation to the very journal the letter appeared in (a non-existent self-reference).
Withdrawn 28 Jan 2026 · source ↗

Court of Appeal of Alberta

Legal
Canada
Legal authorities
CAD 17,550 personal costs
A factum drafted with generative AI cited non-existent authorities — nearly half the table of authorities did not exist.
Sanctioned 22 Jan 2026 · source ↗

Bombay High Court

Legal
India
Legal authorities
Fake judgment; Rs 50,000 cost order
Submissions cited a non-existent judgment that the court and its law clerks could not trace despite extensive verification.
Sanctioned 7 Jan 2026 · source ↗

World Bank Group

Government
United States
References & studies
At least 14 fake references
At least 14 in-text references appear non-existent, a pattern characteristic of LLM-generated citations.
Withdrawn 19 Dec 2025 · source ↗

US District Court, District of Oregon (Magistrate Judge Mark D. Clarke)

Legal
United States
Legal authorities +1
$110,000 penalty
15 references to non-existent cases and 8 fabricated quotations; one attorney then attempted a cover-up by removing the false material and refiling.
Sanctioned 12 Dec 2025 · source ↗

Journal of Academic Ethics (Springer Nature)

Academic
Ethiopia
References & studies
19 of 29 references fabricated
At least 19 of 29 references appear fabricated; 18 Google Scholar links return empty; author admitted ChatGPT-generated references.
Withdrawn 2 Dec 2025 · source ↗

Deloitte

Consulting
Canada
References & studies
CA$1.6M report, fake citations
At least 4 non-existent academic citations; Deloitte said AI wrote 'a limited number of erroneous citations' and stands behind the report.
Corrected 25 Nov 2025 · source ↗

DARU Journal of Pharmaceutical Sciences (Springer Nature)

Academic
Iran
References & studies
5 of 17 references invented
5 of 17 references do not exist and 3 more have wrong DOIs; one fabricated cite credited Retraction Watch's co-founder with a non-existent article.
Corrected 13 Nov 2025 · source ↗

EY (Ernst & Young) Canada

Consulting
Canada
References & studies +1
16 of 27 sources fabricated
16 of 27 sources fabricated, misattributed or broken, including an invented McKinsey statistic.
Withdrawn Nov 2025 · source ↗

US District Court, Southern District of Indiana

Legal
United States
Legal authorities
Fake cites in 3 briefs
Fake AI-generated citations across three separate briefs.
Sanctioned Nov 2025 · source ↗

Journal of the Law Society of Scotland

Media
United Kingdom
Sources & experts +1
Feature pulled over phantom sources and fabricated quotes
Quotes falsely attributed to interviewees the reporter could not produce. On the balance of the evidence the editor concluded the quotations were "disputed and otherwise problematic," falsely attributed, and likely fabricated, in breach of the Journal's editorial guidelines and the author's contractual obligations.
Withdrawn 31 Oct 2025 · source ↗

Texas state district court (Travis County)

Legal
United States
Legal authorities +1
6+ misquoted/bogus cites
At least six cited cases said something different from what the brief claimed; quotes blamed on AI.
Sanctioned 22 Oct 2025 · source ↗

Deloitte

Consulting
Australia
References & studies +1
A$439,000 report, partly refunded
Fabricated academic references plus a quote misattributed to a Federal Court judge.
Refunded 6 Oct 2025 · source ↗

ENISA (EU Agency for Cybersecurity)

Government
European Union
References & studies
26 of 492 footnotes wrong
AI-hallucinated source references: 26 of 492 footnotes were incorrect, with many dead source links; AI had made 'editorial revisions'.
Corrected Oct 2025 · source ↗

California Court of Appeal, Second Appellate District

Legal
United States
Legal authorities +1
21 of 23 quotes made up
21 of 23 quotes from cited cases were fabricated; briefs drafted with ChatGPT, Claude, Gemini and Grok.
Sanctioned 12 Sep 2025 · source ↗

General Division of the High Court of Singapore (Assistant Registrar Tan Yu Qing)

Legal
Singapore
Legal authorities
S$800 personal costs — Singapore's first AI-citation sanction
Written submissions filed 1 June 2025 cited a case — fabricated name and citation, generated by a generative-AI tool used by a junior lawyer — for the proposition that ad hoc loans between acquaintances are not moneylending. The authority was not in the bundle and could not be traced. Counsel first described it as a clerical error and admitted the citation was fictitious, and that AI had been used, only when pressed by the court.
Sanctioned Sep 2025 · source ↗

Business Insider

Media
United States
Sources & experts +1
40+ essays pulled
Essays whose named subjects, anecdotes and quoted details could not be verified — part of a fabricated-contributor network, with details contradicting between pieces.
Withdrawn Sep 2025 · source ↗

Government of Newfoundland & Labrador — Education Accord NL

Government
Canada
References & studies
15+ fake sources in a 10-year plan
At least 15 citations to non-existent journal articles, including a reference to a 2008 NFB film the NFB confirms does not exist (traced to a style guide's fictitious examples).
Corrected 28 Aug 2025 · source ↗

US District Court, District of Arizona

Legal
United States
Legal authorities +1
~12 of 19 cites fabricated
Of 19 citations, only 5-7 existed; opinions were falsely attributed to three real Arizona federal judges.
Sanctioned 14 Aug 2025 · source ↗

Sandiganbayan (Philippines anti-graft court)

Legal
Philippines
Legal authorities +1
Non-existent authorities across 3 motions
Three AI-assisted motions cited authorities that were non-existent or did not contain the quoted passages, with incorrect docket numbers and dates.
Sanctioned 7 Aug 2025 · source ↗

Wired (Condé Nast)

Media
United States
Sources & experts +1
Phantom expert, no fact-check
Quotes from sources who do not appear to exist (e.g. 'Jessica Hu, an ordained officiant in Chicago'), supplied by a fictitious freelancer whose copy appears AI-generated.
Withdrawn Aug 2025 · source ↗

US District Court, Northern District of Alabama

Legal
United States
Legal authorities
Fake cites; firm disqualified
Motions contained false case citations the attorneys confirmed were ChatGPT hallucinations.
Sanctioned 23 Jul 2025 · source ↗

Springer Nature

Academic
United Kingdom
References & studies
25 of 46 references unverifiable
Publisher could not verify 25 of 46 references; of 18 checked, about two-thirds did not exist or had substantial errors.
Withdrawn 16 Jul 2025 · source ↗

England & Wales High Court (Dame Victoria Sharp, President of the King's Bench Division)

Legal
United Kingdom
Legal authorities
18 of 45 citations fabricated
18 of 45 citations were fabricated.
Sanctioned 6 Jun 2025 · source ↗

White House / US MAHA Commission

Government
United States
References & studies
7+ nonexistent studies
At least seven cited studies were non-existent, with dozens of further reference errors; many footnotes carried the 'oaicite' (ChatGPT) marker.
Corrected 22 May 2025 · source ↗

Chicago Sun-Times & Philadelphia Inquirer

Media
United States
Sources & experts +1
10 of 15 books didn't exist
10 of 15 recommended books were entirely invented; other items quoted fake or misidentified experts.
Withdrawn 15 May 2025 · source ↗

US District Court, Central District of California

Legal
United States
Legal authorities
~9 of 27 cites wrong
About 9 of 27 citations were wrong and at least 2 cited authorities did not exist; AI generated the first draft.
Sanctioned 6 May 2025 · source ↗

High Court of Justice (King's Bench Division)

Legal
United Kingdom
Legal authorities
5 fake cases
Grounds for judicial review cited five non-existent authorities, including a fabricated Court of Appeal decision.
Sanctioned 3 Apr 2025 · source ↗

US District Court, District of Wyoming

Legal
United States
Legal authorities
8 of 9 cases fake
8 of 9 cases cited in the motions did not exist — generated by the firm's in-house AI tool 'MX2.law'.
Sanctioned 24 Feb 2025 · source ↗

Supreme Court of Israel

Legal
Israel
Legal authorities +1
36 fake citations
A pleading relied on 36 AI-generated purported judgments: 5 do not exist, 14 had no matching case, and 24 included unconnected quotes.
Sanctioned 20 Feb 2025 · source ↗

Tribunal de Justiça de Santa Catarina (TJSC)

Legal
Brazil
Legal authorities
Fabricated precedents; 10% bad-faith fine
Cited case-law precedents and legal doctrine that were wrong or wholly fictitious; the lawyer admitted 'inadvertent' ChatGPT use.
Sanctioned 18 Feb 2025 · source ↗

US District Court, District of Minnesota

Academic
United States
References & studies
2 of 15 citations fabricated
Of 15 citations in a misinformation expert's declaration, at least two referenced academic articles that do not exist — admitted ChatGPT hallucinations.
Sanctioned 10 Jan 2025 · source ↗

KwaZulu-Natal High Court, Pietermaritzburg

Legal
South Africa
Legal authorities
7 of 9 cited cases did not exist
Of nine cited cases, only two were found to actually exist; the fabricated authorities were most likely AI-generated.
Sanctioned 8 Jan 2025 · source ↗

US Bankruptcy Court, Northern District of Illinois

Legal
United States
Legal authorities
Rule 9011 AI sanctions
Portions of the briefs were prepared with AI and contained fabricated citations, violating Bankruptcy Rule 9011.
Sanctioned 2025 · source ↗

Court of Appeal of New Zealand

Legal
New Zealand
Legal authorities
AI-cited non-existent authorities
A self-represented litigant's submissions relied on non-existent case authorities; the court noted the material appeared AI-drafted.
Withdrawn 21 Nov 2024 · source ↗

Federal Circuit and Family Court of Australia

Legal
Australia
Legal authorities
First Aussie lawyer regulated over AI fakes
An AI-generated list of authorities cited cases that did not exist, with inaccurate summaries; none verified before filing.
Sanctioned Oct 2024 · source ↗

Cody Enterprise (Wyoming)

Media
United States
Quotes
7 stories, fake quotes
Seven stories contained AI-generated fabricated quotes from six people, including invented quotes attributed to Wyoming's governor.
Corrected Aug 2024 · source ↗

KPMG (Australian government report)

Government
Australia
References & studies
Invented a Retraction Watch reference
A fabricated reference attributed to Retraction Watch's own founders for a paper they never wrote.
Corrected 4 Mar 2024 · source ↗

BC Supreme Court (Canada)

Legal
Canada
Legal authorities
Two fabricated cases; lawyer pays costs
An application cited two ChatGPT-hallucinated cases that do not exist in any record.
Sanctioned 20 Feb 2024 · source ↗

US Second Circuit

Legal
United States
Legal authorities
Fake case invented by ChatGPT
Cited a non-existent decision (Matter of Bourguignon v. Coordinated Behavioral Health Servs.) that the attorney admitted was generated by ChatGPT.
Sanctioned 30 Jan 2024 · source ↗

US District Court, E.D. Texas

Legal
United States
Legal authorities +1
$2,000 penalty + AI ethics course
Case citations to non-existent cases and fabricated quotations generated by AI.
Sanctioned 2024 · source ↗

Parliament of Australia — Senate consulting-services inquiry

Government
Australia
Claims & data
Bard-fabricated allegations vs the Big Four
Multiple fictitious case studies of Big Four wrongdoing generated by Google Bard (false NAB, Probuild, Patisserie Valerie, Westpac claims) — none occurred.
Corrected 6 Nov 2023 · source ↗

US District Court, Southern District of New York

Legal
United States
Legal authorities
The foundational case (2023)
The original 2023 fake-citation case — ChatGPT-invented cases cited in a federal filing.
Sanctioned 22 Jun 2023 · source ↗

Die Aktuelle (Funke Mediengruppe)

Media
Germany
Quotes
€200,000 fake interview
An entire 'interview' of fabricated, AI-generated quotes attributed to Michael Schumacher — he gave no such interview.
Sanctioned 22 Apr 2023 · source ↗
No incidents match those filters.
What the data shows

Four things the register makes plain.

Not a list of anecdotes — a pattern. These are read straight off the verified incidents below and update as the register grows.

3 → 31

The pace is rising.

Logged incidents climbed from 3 in 2023 to 31 in 2025, with 31 already logged in the first 8 months of 2026. The independent rate agrees: the Lancet finds fabricated references up a 12-fold rise in two years.

31

It is not just lawyers.

The courtroom cases are the famous ones — but beyond the 41 of those, 31 more incidents hit the Big Four (KPMG, Deloitte, EY), national governments, peer-reviewed journals and major newsrooms, across 22 countries. No corner of professional writing is exempt.

39

The consequences are real.

39 of 72 incidents ended in a formal sanction and 20 documents were withdrawn outright. Penalties run up to a $110,000 court fine, suspended law licences and struck filings.

1 pattern

It looks exactly right.

Every case is the same defect: a confident, plausible citation, source or quote that a professional read, trusted and signed off — and that no human reviewer caught until after it was published.

Where Ukweli comes in

That last pattern is the whole point. Every entry in this register began as a reference that passed human review and failed only in public. That single class of defect — fabricated or unverifiable citations, sources, quotes and claims — is exactly what Ukweli is built to catch before a document leaves your hands. It reads every citation and claim, not just the suspicious ones, and tells you plainly whether to hold, review or clear. We keep this register because it is the public evidence for the problem Ukweli works on.

How this is built

Verification is the product.

The rule

Every incident is confirmed against at least one reputable source — mainstream news, a legal or official record, or the organisation itself. If it cannot be verified, it is not listed. An empty week is an honest week.

What counts

Documents withdrawn, refunded, corrected, retracted or sanctioned specifically because of AI-generated, fabricated or hallucinated citations or sources. Document-authenticity and version-integrity disputes are tracked separately, not folded in.

For researchers, journalists and analysts

Use this data.

The full register is published as a single JSON file under a Creative Commons CC BY 4.0 licence. You can reuse it commercially, redistribute it and build on it — including for machine-readable and AI use — provided you attribute it. There is no key, no sign-up and no rate limit.

The endpoint

https://ukweli.io/integrity/integrity-watch-dataset.json

A single JSON document: a meta block and an incidents array. It is rewritten every time the register is published, so the URL is stable and the contents are current. meta.last_updated and meta.record_count tell you which version you have.

What is in a record

Each incident carries what happened (org, country, sector, doc_type, what_fabricated), what it cost (consequence_type, consequence_detail), when (incident_date with an explicit date_precision, plus date_logged), and the receipts (source_urls).

Every field is documented inside the file itself, in meta.fields — read it there rather than from any summary, including this one. meta.verification_rule states the standard every entry had to meet.

Two things to read before you use it

Dates carry their own precision. Where a source gave only a month or a year, date_precision says so. Treat incident_date as accurate only to that precision — do not render it as a day.

This is a register, not a survey. It counts incidents that were reported and could be verified, so it is a floor and not a rate. It cannot tell you what share of all documents contain fabrications, and nothing here should be presented as though it can.

How to attribute it

Use this line, or something close to it:

“Ukweli Integrity Watch, The Orkestra, CC BY 4.0 — ukweli.io/integrity”

Please link to ukweli.io/integrity, and cite the individual incident page where you are relying on a single record — each one has its own permanent URL and lists its sources in full. If you publish something built on this data, we would like to see it: hello@ukweli.io.

The two-second test
Could you spot the fabricated citation?

Three citations pulled from real documents. One checks out. One cites a law that no longer exists. One was simply made up. Reveal what Ukweli sees in each.

“… in accordance with the Stock Exchange Control Act, 1985 (as amended).”
From a corporate governance document
“Fabricated citations: an audit across 2.5 million biomedical papers. The Lancet (2026).”
From a research summary
“Case study: Swiss Federal Railways runs AI agents that plan, book and optimise passenger journeys in real time.”
From a Big-Four thought-leadership report
Most readers can’t tell — the fabrications look exactly like the real ones. Ukweli reads every citation, not just the suspicious ones. Integrify your document →
Before you publish

Don't become an entry.

Ukweli checks a document's claims and citations before it leaves your hands — and tells you plainly whether to hold, review, or clear it. Ukweli it first.