You ask ChatGPT a question. It gives you a confident, detailed answer with specific numbers and a citation from a research paper.

There’s just one problem: the paper doesn’t exist. The numbers are invented. The entire answer sounds perfect — and is completely made up.

This is called an AI hallucination. And it happens far more often than most people realize.

This guide explains what AI hallucinations are, why they happen, how often each major model hallucinates, and — most importantly — how to catch them before they cause real damage.

What Is an AI Hallucination?

An AI hallucination is when an AI tool generates information that sounds convincing but is factually wrong, fabricated, or completely made up.

The term “hallucination” comes from the idea that the AI is “seeing” things that aren’t there — like a person hallucinating. The AI isn’t lying on purpose. It’s producing what statistically sounds right based on patterns in its training data, without any mechanism to check whether it’s actually true.

Here’s what hallucinations look like in practice:

  • Fake citations: AI invents a research paper with a real-sounding title, real author names, and a plausible journal — but the paper doesn’t exist.
  • Invented statistics: “According to a 2025 study, 73% of remote workers prefer hybrid schedules.” Sounds convincing. Completely fabricated.
  • Wrong facts stated confidently: “The Eiffel Tower was completed in 1901.” (It was 1889.) No hedging, no uncertainty — just wrong.
  • Fictional events: AI describes a historical event, complete with dates and details, that never happened.
  • Fabricated people: AI creates biographical details about real people that are entirely made up.

The dangerous part isn’t that AI gets things wrong. It’s that it gets things wrong with total confidence. Research shows AI models use 34% more confident language — words like “definitely” and “certainly” — when hallucinating compared to when they’re stating actual facts.

Why Does AI Make Things Up?

This is the question that matters most. Understanding why AI hallucinates helps you predict when it’s likely to happen — and protect yourself.

Reason 1: AI Predicts Words, Not Truth

This is the fundamental issue. AI models like ChatGPT, Claude, and Gemini are language models. They predict the most statistically likely next word based on patterns in their training data.

They don’t “know” things the way you know things. They don’t have a database of verified facts they look up. They generate text that sounds like the kind of text that would follow your question — based on billions of examples from the internet.

Sometimes what sounds right IS right. Often, it isn’t.

Think of it like an extremely well-read person who has memorized the rhythm and style of expert writing — but hasn’t actually verified any of the facts. They can produce something that reads like a scientific paper without any of it being true.

Reason 2: Training Rewards Guessing Over Honesty

OpenAI’s own researchers published a paper in 2025 explaining why this problem is so persistent. Their finding was striking: AI is trained to guess, not to admit uncertainty.

They examined the 10 most popular AI benchmarks — the tests used to evaluate how “smart” a model is. Nine out of 10 use binary scoring: correct answer gets a 1, wrong answer or “I don’t know” both get a 0.

This means a model that confidently guesses wrong scores the same as a model that honestly says “I’m not sure.” But a model that guesses right sometimes will beat the honest model on average. So AI learns to always guess — even when it doesn’t know.

As the researchers put it: the system rewards a “fake-it-till-you-make-it” approach over genuine caution.

Reason 3: Training Data Has Gaps

No AI model has been trained on all human knowledge. OpenAI’s researchers introduced the concept of the “singleton rate” — the percentage of facts that appear only once in training data. If 20% of a particular type of fact appears only once in the training data, the model will hallucinate roughly 20% of the time on those types of questions.

Some facts are well-represented in training data (the capital of France, basic math). Others appear rarely or not at all (niche legal precedents, obscure historical events, recent developments). When AI encounters a question about something it has limited data on, it fills in the gaps with plausible-sounding fiction.

Reason 4: It Can’t Say “I Don’t Know” Naturally

Humans naturally say “I’m not sure” or “let me check.” Language models don’t have this instinct built in. They’re optimized to produce complete, fluent responses. Saying “I don’t know” is a skill that has to be specifically trained into them — and most models aren’t trained on it enough.

Some newer models are getting better at expressing uncertainty, but it’s still not reliable. Even when a model says “I think” or “I believe,” it doesn’t necessarily mean it’s less confident — those are just common patterns in English text.

How Often Does AI Hallucinate? The Real Numbers

Hallucination rates vary dramatically by model, task, and how you measure them. Here are the actual benchmarked numbers from 2025-2026.

Summarization Tasks (Vectara Benchmark)

This measures how often models make things up when summarizing a document they were just given — the easiest possible task.

ModelHallucination Rate
Google Gemini 2.0 Flash0.7%
OpenAI o3-mini0.8%
Google Gemini 2.0 Pro0.8%
OpenAI GPT-4.5 Preview1.2%
OpenAI GPT-4o1.5%
OpenAI GPT-41.8%
Claude 3.7 Sonnet4.4%
Meta Llama 4 Maverick4.6%
Claude 3 Opus10.1%

Those numbers look low — but remember, this is the easiest test. The model has the source document right in front of it and still makes things up.

By Domain (Where It Gets Scary)

Hallucination rates explode when you move into specialized topics:

DomainBest ModelsAll Models Average
General Knowledge0.8%9.2%
Historical Facts1.7%11.3%
Technical Documentation2.9%12.4%
Financial Data2.1%13.8%
Medical/Healthcare4.3%15.6%
Scientific Research3.7%16.9%
Coding & Programming5.2%17.8%
Legal Information6.4%18.7%

Legal questions have the highest hallucination rate at 18.7%. That’s nearly one in five answers containing fabricated information. Medical and scientific queries aren’t far behind. These are exactly the domains where wrong information is most dangerous.

The Good News: It’s Improving

The best-performing models have improved from a 21.8% hallucination rate in 2021 to 0.7% in 2025 — a 96% reduction in four years. There are now four models with sub-1% hallucination rates. The industry has invested over $12.8 billion specifically in hallucination reduction between 2023 and 2025.

But “improving” doesn’t mean “solved.” Even at 0.7%, if you ask an AI 1,000 questions, roughly 7 answers will contain fabricated information — with no warning label.

Real Examples That Actually Happened

These aren’t hypothetical. These are documented cases of AI hallucinations with real consequences.

The Lawyers Who Cited Fake Cases

In 2023, a New York lawyer used ChatGPT to research a legal brief. The AI generated six case citations that looked completely legitimate — correct formatting, plausible case names, real-sounding legal reasoning. None of the cases existed. The lawyer was fined $5,000.

Since then, the problem has gotten worse, not better. By 2025, courts were seeing AI hallucination cases at a rate of two to three per day, up from two per week before spring 2025. A California judge fined two law firms $31,000 for a brief where 9 out of 27 citations were fabricated. A federal judge fined attorneys representing MyPillow CEO Mike Lindell $3,000 each for AI-generated fake citations.

An entire database now tracks these cases — over 1,030 documented legal AI hallucinations and counting.

Google’s Glue-on-Pizza Disaster

When Google launched AI Overviews in search results in 2024, its AI confidently told users to add non-toxic glue to pizza sauce to keep cheese from sliding off. The source? An 11-year-old joke comment from a Reddit user. The AI also suggested eating rocks for their mineral content and mixing bleach with vinegar (which produces toxic chlorine gas).

Google’s AI couldn’t distinguish between authoritative health information and internet jokes.

Medical Misinformation

Researchers at Flinders University found that leading AI models could be manipulated to produce false health claims complete with fabricated citations from real medical journals. The papers didn’t exist, but the citations looked real enough that even medical professionals would need to manually verify each one.

The $67.4 Billion Problem

According to industry estimates, global business losses attributed to AI hallucinations reached $67.4 billion in 2024. Employees now spend an average of 4.3 hours per week just verifying whether AI output is accurate — roughly $14,200 per employee per year in pure verification overhead.

How to Catch AI Hallucinations: A Practical Guide

You don’t need technical skills to spot hallucinations. Here are concrete techniques that work.

1. Watch for the Red Flags

Research has identified patterns in how AI hallucinates:

  • Round numbers: Fabricated statistics end in 0 or 5 about 3.7 times more often than real statistics. If every number the AI gives you is suspiciously round, be skeptical.
  • Overly confident language: When AI uses words like “definitely,” “certainly,” or “it’s well-established that…” it’s actually 34% more likely to be hallucinating.
  • Too-perfect answers: If the AI’s response fits your question suspiciously well — every point perfectly supports your argument, every stat aligns — it may be telling you what you want to hear rather than what’s true.
  • Specific citations you haven’t seen before: If the AI cites a specific paper, book, or study, Google the exact title. If it doesn’t exist, that’s a hallucination.

2. Use the “Google Test”

The simplest hallucination detection method: take any specific claim the AI makes — a statistic, a date, a name, a quote — and search for it. If you can’t find it from any other source, the AI likely invented it.

This takes 10 seconds per claim and catches the majority of hallucinations.

3. Ask the AI to Show Its Work

When you need accurate information, structure your prompts to reduce hallucinations:

Instead of: “What are the statistics on remote work productivity?”

Try: “What are the statistics on remote work productivity? Only include data from studies you’re confident are real. For each statistic, provide the source name and year. If you’re not sure about something, say so rather than guessing.”

This won’t eliminate hallucinations, but research shows that asking “Are you sure about this?” or requesting sources reduces hallucination rates by roughly 17%.

4. Break Complex Questions Into Steps

AI is more likely to hallucinate on complex, multi-part questions. Instead of asking one big question, break it into smaller, verifiable pieces.

Instead of: “Give me a complete analysis of the electric vehicle market including market size, growth rate, top companies, and regulatory environment.”

Try asking each part separately, verifying each answer before moving on. This is called prompt chaining, and it significantly reduces compounding errors.

5. Cross-Check Between Models

If accuracy matters, ask the same question to multiple AI tools. If ChatGPT and Claude give you the same specific statistic, it’s more likely to be real. If they give you different numbers, at least one is hallucinating — and possibly both.

6. Never Trust AI for High-Stakes Decisions Without Verification

This is the most important rule. If the information will be used for:

  • Legal filings or contracts
  • Medical decisions
  • Financial decisions
  • Published content with your name on it
  • Anything where being wrong has real consequences

Always verify against primary sources. AI is a research assistant, not an authority. If you’re using AI at work, this habit is non-negotiable.

What’s Being Done to Fix This

The AI industry isn’t ignoring the problem. Here’s what’s happening.

Retrieval-Augmented Generation (RAG)

Think of RAG as giving AI an open-book exam instead of a closed-book exam. Instead of relying only on what it memorized during training, the AI retrieves relevant documents at the time of your question and bases its answer on actual source material.

RAG reduces hallucinations by up to 71% in studies. Many enterprise AI products already use this approach. When you use tools like Perplexity AI or Microsoft Copilot with your documents, you’re using a form of RAG.

Better Benchmarks

Following OpenAI’s research, the industry is starting to build benchmarks that reward saying “I don’t know.” If models are evaluated on honesty rather than just confidence, they’ll learn to be more cautious. This is a slow shift, but it’s happening.

Self-Consistency Checking

Google has developed methods where the AI generates multiple answers to the same question and checks them against each other. If the answers disagree, the model flags uncertainty. This technique reduces hallucinations by up to 65%.

Human-in-the-Loop Systems

Many AI products now include verification layers where AI agents — autonomous AI systems that take actions — pause and check with humans before acting on potentially hallucinated information. This is especially important for high-stakes applications in law, medicine, and finance.

Model Size Matters

Larger models generally hallucinate less:

  • Models under 7 billion parameters: 15-30% average hallucination rate
  • Models 7-70 billion parameters: 5-15%
  • Models over 70 billion parameters: 1-5%

This is one reason why the most capable models from OpenAI, Google, and Anthropic perform better — they’re significantly larger than open-source alternatives.

How AI Hallucinations Affect You

Even if you’re not a developer or a lawyer, hallucinations affect anyone who uses AI tools.

If you use AI for research: Every statistic, citation, and factual claim needs verification. AI is excellent for understanding concepts and generating ideas — but treat specific facts as unverified until you confirm them.

If you use AI for writing: AI-generated content may contain plausible-sounding claims that are fabricated. If you publish them, your credibility is on the line — not the AI’s.

If you use AI for learning: AI can explain concepts brilliantly. But it can also teach you things that are subtly wrong. Use AI to understand how things work, then verify specific facts through textbooks, official documentation, or peer-reviewed sources.

If you rely on AI search results: Google’s AI Overviews, Perplexity, and other AI search tools can hallucinate just like chatbots. Don’t assume AI-generated search summaries are more accurate than the underlying sources — click through and verify.

The underlying skill here is the same one we discussed in our guide on prompt engineering: the better you communicate with AI, the better your results. Specific, well-structured prompts with clear constraints produce fewer hallucinations than vague, open-ended questions.

The Bottom Line

AI hallucinations are when AI tools confidently make things up. It happens because these models predict what sounds right rather than what is right — and they’re trained to guess rather than admit uncertainty.

The best models have gotten dramatically better, dropping from 21.8% hallucination rates in 2021 to 0.7% in 2025. But no model is hallucination-free, and rates spike sharply on legal, medical, and scientific questions — exactly the domains where accuracy matters most.

The fix isn’t to stop using AI. It’s to use it with your eyes open. Treat AI like a brilliant but unreliable research assistant: great for first drafts, explanations, and brainstorming — but never the final word on facts. Verify anything that matters. And remember: the most confident-sounding answer isn’t always the most accurate one.

Sources and Further Reading

This article draws on peer-reviewed research, official documentation, and verified industry data: