AI Hallucinations: The Crisis Nobody Solved

1,006 court cases involving AI-hallucinated content worldwide

A legal researcher named Damien Charlotin maintains a database that tracks every court case worldwide where a party relied on AI-hallucinated content. As of early March 2026, that number has crossed 1,006. More than a thousand times, someone trusted an AI's confident-sounding answer, put it in a legal filing, and got caught. 90% of those cases were written in 2025 alone.

This isn't a fringe problem. In the United States alone, 518 cases have been documented since the start of 2025, implicating 128 lawyers, including attorneys from Am Law 100 firms. And that's just the legal profession, where hallucinations leave a paper trail. In medicine, finance, consulting, and everyday enterprise use, the problem is harder to count but no less real.

What Hallucination Actually Means

When an AI hallucinates, it generates information that sounds authoritative but is fabricated. Not a rough approximation. Not an honest mistake. A confident, detailed, citation-complete invention. It will give you a court case that never happened, a study that was never published, a statistic that no one ever measured. And it will do so in a tone indistinguishable from its accurate responses.

Duke University published an analysis in January 2026 asking why this is still happening. Their answer boils down to four problems: models are trained on internet data that mixes fact with fiction, they're optimized to sound helpful rather than accurate, benchmark systems reward confident answers over honest uncertainty, and they fundamentally lack the ability to know what they don't know. AI sycophancy, the tendency to validate whatever the user seems to want, makes it worse.

The Courtroom Becomes a Case Study

The most famous case is Mata v. Avianca from 2023, where two New York attorneys submitted a brief containing six entirely fabricated court decisions generated by ChatGPT. Judge Castel imposed a $5,000 fine and the case became a cautionary tale. But the cautionary tale didn't caution anyone.

In February 2025, three attorneys from Morgan and Morgan, the 42nd largest U.S. law firm, filed motions in a products liability case against Walmart. They used the firm's own AI platform to add case law. Of nine cases cited, eight were nonexistent. The attorneys were fined and one had their admission revoked.

In July 2025, two lawyers representing MyPillow CEO Mike Lindell submitted an AI-generated filing with over two dozen errors, including fabricated cases. When caught, one attorney claimed it was a "draft" filed by accident. The "final" version contained additional errors. Each was fined $3,000.

An Oregon attorney was fined $15,500 in December 2024 for citing fake AI-generated cases. The fine was elevated because, according to the court, they were not "adequately forthcoming, candid, or apologetic." The pattern is consistent: AI generates convincing fictions, professionals trust them without checking, and the consequences scale with the stakes.

Beyond the Courtroom

In medicine, researchers found that OpenAI's Whisper transcription tool, used by approximately 30,000 clinicians and 40 health systems, fabricates text that was never spoken. About 1% of transcriptions contained hallucinations, including invented medications like "hyperactivated antibiotics" and inappropriate racial commentary injected into patient records. A Cornell University study estimated 40% of Whisper's hallucinations could have harmful consequences for patients.

In consulting, Deloitte produced a $290,000 report for the Australian government that contained fabricated academic references and a false court quote. A University of Sydney researcher identified the hallucinated citations. Deloitte issued a partial refund. Then it happened again: a $1.6 million Deloitte report for a Canadian provincial government was found to contain similar AI-generated errors in November 2025.

In consumer search, Google's AI Overviews launched in May 2024 and immediately recommended adding glue to pizza sauce, eating rocks daily, and smoking while pregnant. Google CEO Sundar Pichai later called hallucinations "an unsolved problem" and "in some ways an inherent feature."

And in the corporate world, Klarna replaced roughly 700 customer service workers with an AI assistant, initially claiming $10 million in savings. Within six months, customer satisfaction dropped 22%. CEO Sebastian Siemiatkowski publicly admitted "we went too far" and began rehiring human agents.

The Numbers That Matter

You'll see a figure circulating online: $67.4 billion in global losses from AI hallucinations. We traced this number back to its source. It comes from AllAboutAI, a content aggregation site that openly admits its statistics "are based on information available from various secondary sources" and "some figures may not be independently verified." No methodology, no sample size, no industry breakdown. It's an estimate of unknown provenance that went viral through circular citation. We won't repeat it as fact, and you should be skeptical of anyone who does.

What IS verifiable from primary research: Stanford's Human-Centered AI Institute published the first peer-reviewed evaluation of commercial legal AI tools in 2025. Their findings: Lexis+ AI hallucinates 17% of the time, Westlaw AI hallucinates 33% of the time, and GPT-4 without retrieval augmentation hallucinates 43% of the time. These are tools that lawyers are using right now to research case law.

The Vectara Hallucination Leaderboard, updated February 25, 2026 and the most comprehensive benchmark available, tests models on grounded summarization. The best-performing models still hallucinate 1.8% of the time. Mainstream models like GPT-5, Claude, and Gemini-3-Pro hallucinate at rates above 10%. The models are getting better. They are nowhere close to reliable.

RAG Is Not the Fix Everyone Thinks

Retrieval-Augmented Generation, the technique of grounding AI responses in specific documents, is widely promoted as the solution to hallucination. And it helps. The Stanford study showed RAG reduced hallucination rates from 43% to 17-33% in legal contexts. A cancer research study found RAG chatbots using reliable sources achieved 0% hallucination with GPT-4 on domain-specific questions.

But RAG doesn't eliminate the problem. It constrains it. When the retrieval works perfectly and the documents are accurate, RAG performs well. When retrieval fails, when the model can't find a relevant passage, or when the question requires reasoning beyond the source material, it falls back to the same pattern: confident fabrication. The Stanford researchers specifically noted that RAG-based tools still "ichallucinate at a meaningful rate" and cautioned against vendor claims that RAG makes AI reliable.

What This Means for You

If you're using AI tools at work, and according to Pew Research roughly 35% of Americans now are, hallucination is a problem you need to understand. Not because the tools aren't worth using. They are. The productivity gains from drafting, brainstorming, and pattern recognition are real. But the same tools that save you hours can also fabricate a citation, invent a statistic, or confidently describe something that never happened. The gap isn't in the technology. It's in how well most people understand what these tools actually do.

The 1,006 court cases in Charlotin's database aren't there because AI is bad at law. They're there because professionals assumed that a confident answer was a correct answer. The Deloitte reports weren't embarrassing because AI can't write. They were embarrassing because nobody checked the citations before submitting to a government client.

Google's Sundar Pichai called hallucination "inherent." Duke University's research explains why it persists. That's not a reason to avoid these tools. It's a reason to use them with a clear understanding of their failure modes. The lawyers who got fined weren't bad at their jobs. They just didn't know what AI doesn't know. That's a gap in understanding, and it's one that's entirely fixable.

Update, March 9, 2026: The Charlotin database count has been updated from 979 to 1,006 to reflect the current live total. All other figures in this post remain unchanged.