mikias\abera

CITATION FAILURES

When AI-fabricated citations ship

—— 1,737 documented incidents: every court case on record across 41 countries, plus the consulting reports, government policies, and research failures that never reach a courtroom. Rebuilt automatically every week, every entry linking its source document.

1,737
documented incidents
122
added in the last 30 days
41
countries
297
cases with monetary penalties

data refreshed 2026-07-08 · court cases: Damien Charlotin's AI Hallucination Cases database (CC0)

THE BOARD

Every documented case, searchable

Courts leave records, so they're where this failure is best documented: a judge finds citations to authority that doesn't exist, quotes no opinion contains, or real cases misread. Each row links the court document. Search it, filter it, or take the whole dataset as JSON.

1,731 cases, newest first · loading the full dataset

Patterson v. Nuvision Credit Union ↗

500 USD · Monetary sanction; Bar Referral (UPL)

2026-07-02 · CA California (4d) · USA · Pro Se Litigant · Unidentified · fabricated ×7 · false quotes ×4 · misrepresented ×1

USA v. Organes ↗

no sanction recorded

2026-06-29 · N.D. Illinois · USA · Pro Se Litigant · fabricated ×1 · false quotes ×1

Morgan v. Jones ↗

no sanction recorded

2026-06-26 · Ohio · USA · Pro Se Litigant · fabricated ×1 · false quotes ×1

Barari v. Burgess ↗

Brief Struck

2026-06-26 · CA Arizona (1d) · USA · Pro Se Litigant · Unidentified · fabricated ×1 · misrepresented ×1

Nady v. Al Sadi, 2026 ONSC 3760 ↗

1 CAD · AI misuse increasing adverse costs award

2026-06-26 · Ontario SCJ · Canada · Pro Se Litigant · Unidentified · fabricated ×1 · false quotes ×1 · misrepresented ×1

Paris v. Paris ↗

no sanction recorded

2026-06-25 · CA Texas · USA · Pro Se Litigant · fabricated ×1

Kim v. Wexel ↗

Warning

2026-06-24 · CA Georgia · USA · Pro Se Litigant · fabricated ×1

Landberg v City of New York ↗

10500 USD · Monetary Sanction

2026-06-23 · CA NY (2d) · USA · Lawyer · Unidentified · fabricated ×3 · misrepresented ×2

THE RATE

2026 is on pace to be the worst year on record

Cases per month since the first documented decision in April 2023. One honest caveat: detection is growing alongside occurrence. Trackers and forensic reviews now catch failures that would've gone unnoticed two years ago, so part of the climb is better looking, not just more failing. The current month renders lighter because it's still in progress.

50100150200Apr 2023: 1 casesMay 2023: 2 casesJun 2023: 3 casesJul 2023: 1 casesAug 2023: 0 casesSep 2023: 2 casesOct 2023: 2 casesNov 2023: 4 casesDec 2023: 1 casesJan 2024: 2 cases2024Feb 2024: 4 casesMar 2024: 4 casesApr 2024: 3 casesMay 2024: 1 casesJun 2024: 1 casesJul 2024: 6 casesAug 2024: 8 casesSep 2024: 5 casesOct 2024: 5 casesNov 2024: 10 casesDec 2024: 10 casesJan 2025: 16 cases2025Feb 2025: 15 casesMar 2025: 24 casesApr 2025: 28 casesMay 2025: 47 casesJun 2025: 44 casesJul 2025: 82 casesAug 2025: 85 casesSep 2025: 90 casesOct 2025: 120 casesNov 2025: 124 casesDec 2025: 152 casesJan 2026: 133 cases2026Feb 2026: 126 casesMar 2026: 172 cases172Apr 2026: 124 casesMay 2026: 147 casesJun 2026: 115 casesJul 2026: 12 cases (month in progress)

BEYOND THE COURTROOM

The failures that never reach a judge

Court cases dominate every tracker because courts publish their findings. The same failure ships inside consulting reports, government policy, and published research, where there's no judge to put it on the record. These entries cover that thinner ground: each one links the record of record, states only what the linked sources say, and attributes every characterization.

KPMG

2026-06 · consulting-report · Global (US-published)withdrawn

Total Experience agentic-AI thought-leadership report. GPTZero's review found only 5 of 45 citations accurate, with the rest fabricated, mangled, or misattributed, and the report was pulled in June 2026. The incident is logged by the OECD AI Incidents Monitor.

primary record ↗ · OECD AIM (db) ↗

last verified 2026-07-08

EY Canada

2026-05 · consulting-report · Canadawithdrawn

Cybersecurity report on loyalty-program fraud, removed from the firm's website. GPTZero's forensic review found 16 of 27 cited sources fabricated, misattributed, or broken, and the report was pulled from EY's website in May 2026, according to trade press coverage.

primary record ↗ · Computing ↗ · Lawyer Monthly ↗

last verified 2026-07-08

Government of South Africa

2026-04 · policy-document · South Africawithdrawn

Draft National Artificial Intelligence Policy. Communications Minister Solly Malatsi announced the draft policy's withdrawal on 26 April 2026 after at least 6 of its 67 bibliography sources were found to be fabricated, according to the government's own statement and independent reporting.

primary record ↗ · Reuters ↗ · News24 ↗ · AIID 1467 (db) ↗

last verified 2026-07-08 · corrections: 2026-07-08: Earlier drafts of this dataset sourced this entry to a single aggregator newsletter. It is now sourced to the government's own withdrawal statement, with independent press and the AI Incident Database entry alongside.

Deloitte Australia

2025-10 · consulting-report · Australiarefund issued

AU$440K assurance review of the Targeted Compliance Framework for the Department of Employment and Workplace Relations. The report contained fabricated academic references and a fabricated Federal Court quote, according to the department and press review. Deloitte repaid the final installment of its contract, and a disclosure of Azure OpenAI use was added only in the revised version.

primary record ↗ · The Guardian ↗ · AFR ↗ · Fortune ↗ · AIID 1193 (db) ↗

last verified 2026-07-08 · corrections: 2026-07-08: Earlier drafts of this dataset cited a refund of about AU$97K from press estimates. No primary source states a dollar amount, so the entry now says final installment only.

Deloitte (Newfoundland and Labrador)

2025-05 · consulting-report · Canada (NL)acknowledged by the org

About C$1.6M Health Human Resources Plan for the Government of Newfoundland and Labrador. Reporting in November 2025 identified at least four citations to non-existent sources in the plan delivered that May, and Deloitte acknowledged selective use of AI in its preparation, according to coverage by The Independent and Fortune.

primary record ↗ · gov.nl.ca release (the plan) ↗ · Fortune ↗ · AIID 1286 (db) ↗

last verified 2026-07-08 · corrections: 2026-07-08: Earlier drafts of this dataset gave the cost as about $1.2M and dated the incident 2026. Independent coverage consistently reports C$1.6M, the plan was delivered in May 2025, and press coverage broke in November 2025.
PubMed-indexed papers with fabricated references
1 in 277
1 in 2,828 in 2023, then 1 in 458 in 2025 · as of first 7 weeks of 2026
source ↗
Incidents logged by the AI Incident Database in 2025
362
up from 233 in 2024 · as of 2025
source ↗
Reproduce the failure type yourself
pick a model in the playground and watch each citation resolve against real registries, or fail
→

METHOD

Fully automated, sources always linked

Two layers, one weekly pipeline. The court-case layer is ingested whole from Damien Charlotin's AI Hallucination Cases database, the definitive tracker of this failure in courtrooms, released CC0 and credited here anyway. It's rebuilt from his published data every week, so corrections upstream propagate automatically.

The cross-domain layer is gathered by an automated weekly sweep with a hard floor: an entry needs the organization's own acknowledgment, correction, withdrawal, or refund; or a court or government record; or at least two independent major outlets each citing a specific named document. Many outlets repeating one origin story count as one source, and such entries wear a visible "reported, unconfirmed" badge until the record hardens. Named organizations of public significance only, never individuals. Incident databases like AIID, the OECD monitor, and AIAAIC are linked as pointers, with thanks, and their entry text is never copied.

Corrections. If anything here is wrong, I want to know. Email mikias@abera.io with the entry, the statement you believe is wrong, and the evidence. For court-case rows, corrections flow through the upstream database and land here on the next weekly rebuild. For cross-domain entries, every change is dated and described on the entry itself. Nothing gets silently edited or deleted, and no organization can pay to be added, softened, or removed.

If you represent a named organization. Every entry is built from public records: your own statements, court or government documents, or independent reporting that names a specific document. If you believe an entry misstates them, email me the specific statement and the evidence. I'll acknowledge within five business days, recheck the entry against its sources, and correct, annotate, or remove anything that doesn't hold up. Named organizations can also send a response, which I'll link from the entry.

DATA

Machine-readable, free, attributed

Court cases (bulk layer): /citations/api/legal-cases.json

Cross-domain incidents: /citations/api/incidents.json

Feeds: RSS · JSON Feed

The scoreboard this page belongs to: mikias.io/citations

Court-case data is CC0 from its source, credited above. The cross-domain compilation is CC BY 4.0: reuse it with attribution to mikias.io. The underlying records belong to their sources, always linked. The incident data itself will never be sold.