Post-Mortem Library

Real AI disasters.
The criteria that catch them.

Every company claims their AI works. We show you exactly how it fails — using public records, court documents, and regulatory filings — then map each failure to the exact review criteria that would have stopped it.

Browse post-mortems
9
AI failures analyzed from 2023–2025 — all 9 published. Sourced from court records, FTC filings, and major news.
0
Customers needed. Every case study cites verifiable public record — no anecdotes, no spin.
Legal
3
Code
1
Medical
1
Financial
2
Image
1
Safety
1

We don't have customer logos. We have disaster post-mortems. Each one shows the exact failure mode, the real-world impact, and the specific review criteria that would have caught it before it shipped.

The library

Nine post-mortems. Zero spin.

Legal Hallucination

When a chatbot invented a refund policy

Air Canada's support bot promised a bereavement fare that didn't exist. The airline lost at tribunal — and set a precedent that companies are liable for what their AI says.

Feb 2024 · Air Canada · 3-min read
Read →
Legal Hallucination

Six fake cases filed in federal court

Two attorneys submitted citations that didn't exist. ChatGPT invented them — complete with fake quotes and docket numbers. $5,000 in sanctions.

Jun 2023 · Mata v. Avianca · 4-min read
Read →
Financial Factual Error

One factual error. $100 billion gone.

Google's Bard demo claimed the James Webb Telescope took the first exoplanet photos. It didn't. The stock dropped 9% in a day. A single fact-check would have caught it.

Feb 2023 · Google · 3-min read
Read →
Legal Illegal Advice

A city chatbot that told businesses to break the law

NYC launched MyCity to help small businesses. Instead, the AI chatbot told them to violate employment law, ignore wage requirements, and skip disability protections.

Jan 2024 · NYC · 4-min read
Read →
Medical Bias

Diverse Nazis and female popes

Google's Gemini image generator produced Black Nazi soldiers, female popes, and refused to generate white people. Google paused the feature globally and Pichai called it "unacceptable."

Feb 2024 · Google · 3-min read
Read →
Financial Math Errors

77 articles. 41 corrections. Trust destroyed.

CNET's AI-written financial articles contained math errors in 53% of pieces — wrong APR calculations, incorrect savings rates, wrong tax formulas. No fact-check process existed.

Jan 2023 · CNET · 3-min read
Read →
Code Security

"Hardcoded secrets" and SQL injection

Amazon Q Developer (Beta) generated code with hardcoded AWS secrets, SQL injection vulnerabilities, and weak cryptographic functions — then labeled it "production ready."

Feb 2024 · Amazon · 4-min read
Read →
Financial False Claims

The "robot lawyer" that couldn't back up its claims

DoNotPay marketed itself as "the world's first robot lawyer" but couldn't substantiate its claims. The FTC settled for $193,000 and banned false AI capability marketing.

Jan 2024 · DoNotPay · 3-min read
Read →
Safety Data Leak

Engineers pasted proprietary code into ChatGPT

Three separate incidents at Samsung: semiconductor source code, internal meeting notes, and production code — all pasted into ChatGPT. Samsung banned the tool company-wide.

Apr 2023 · Samsung · 3-min read
Read →
Our method

How we build each post-mortem

01

Source the failure

Every case starts with public record: court filings, FTC settlements, regulatory actions, or major news reports. We cite primary sources, not secondhand summaries.

02

Reconstruct the output

We reproduce what the AI actually generated — the chatbot response, the fabricated citations, the biased images — and annotate exactly where it went wrong.

03

Map to review criteria

Each failure maps to specific, copy-pasteable review criteria. You can add them to your workflow in the sample builder and test them against your own AI output.

50% off your first $10 — no commitment Get 50% off