Documented AI Failures

A$440K welfare report contained fabricated academic references, non-existent research papers, and a fake judge quote. Used Azure OpenAI GPT-4o.

Second Deloitte incident in 2 months. CA$1.6M healthcare report with at least 4 fake citations, including fictional co-authored papers by researchers who never worked together.

AI summaries falsely claimed Luigi Mangione shot himself, Rafael Nadal came out as gay, and a darts player won before his match even started.

AI customer support bot falsely told users they couldn’t use software on multiple computers, directly causing account cancellations and revenue loss.

Meta’s AI chatbot falsely called a public figure a Holocaust denier and January 6 rioter when asked about him by other users.

AI hiring platform deployed with default credentials ‘123456/123456’. 64 million job applications and chat transcripts left exposed to anyone.

AI company settled for approximately $1.5 billion after training on pirated book libraries — roughly $3,000 per book.

Class action alleging hundreds of thousands of copyrighted books were used without authorization to train enterprise AI agents.

OpenAI’s own internal testing reveals their flagship o3 model answers general knowledge questions incorrectly 51% of the time on SimpleQA benchmark.

Smaller, faster model produced wrong answers 79% of the time on the same SimpleQA benchmark — nearly 4 out of 5 answers are fabricated.

Columbia University’s Tow Center for Digital Journalism found xAI’s Grok 3 generated incorrect news citations 94% of the time.

Humanity’s Last Exam — the hardest AI benchmark — shows even the best model (Gemini 3 Pro) answers only 37.5% of expert-level questions correctly. Models show 34–89% calibration error, meaning they're confidently wrong.

GPT-4o shows 89% calibration error — meaning when it says it’s confident, it’s almost maximally wrong about its own certainty.

Even on simple news summarization tasks, DeepSeek R1 fabricates content 14.3% of the time — inventing facts that aren’t in the source article.

Cross-study analysis shows large language models can hallucinate up to 45% of the time depending on task complexity and domain.

The world’s most important AI benchmark uses OpenAI’s o3-mini — from a model family with 51% hallucination rates — as the judge grading every model’s answers. Meanwhile, 29% of HLE’s own answer key contradicts peer-reviewed literature. Investment decisions, safety claims, and regulatory policy are all informed by scores graded by an unreliable AI against a flawed answer key.