AI Hallucinations: Why AI Confidently Makes Things Up
AI hallucination is when a model states something incorrect with complete confidence. It doesn’t say “I think” or “I’m not sure” — it presents fabricated information as fact. This is the single biggest reliability problem in AI today.
What Hallucination Looks Like
Ask ChatGPT for a list of academic papers on a niche topic. There’s a good chance it invents papers that don’t exist — complete with plausible author names, realistic journal titles, and made-up DOIs. The citations look perfect. They’re completely fabricated.
Ask Claude for a legal precedent. It might cite a real-sounding case name with a fake ruling. Ask any model for statistics and it may generate numbers that feel accurate but came from nowhere.
This isn’t a bug in the traditional sense. It’s a consequence of how these models work.
ELI5: AI Hallucination — Imagine a student who never says “I don’t know.” When they don’t know the answer, they make something up that sounds like it could be right. They’re not trying to lie — their brain is just pattern-matching to produce an answer-shaped response. AI models do the same thing. They’re trained to produce plausible-sounding text, not to be accurate. When they don’t know something, they generate what “sounds right” instead of admitting uncertainty.
Why Models Hallucinate
AI language models are next-token prediction machines. Given a sequence of text, they predict the most likely next word. They’re not looking up facts in a database — they’re generating text that statistically resembles their training data.
When you ask “What year was the first iPhone released?”, the model doesn’t search a fact table. It generates the token “2007” because, in its training data, that number appeared frequently near the words “first iPhone released.” It happens to be correct — but the mechanism isn’t fact retrieval. It’s pattern matching.
When the patterns in training data are clear and consistent, the model gets it right. When they’re ambiguous, sparse, or absent, the model generates what looks right. That’s hallucination.
Key insight: Models don’t know what they don’t know. They have no internal confidence meter that says “I’m making this up.” The generation process is the same whether the output is accurate or fabricated.
How Often Does It Happen?
Hallucination rates vary by model, task, and domain:
| Model | Hallucination Rate (factual Q&A) | Source |
|---|---|---|
| Claude Opus 4 | ~3-5% | Internal testing |
| GPT-4o | ~4-7% | Third-party benchmarks |
| Gemini 2.0 Pro | ~5-8% | Third-party benchmarks |
| Llama 4 (70B) | ~8-12% | Community testing |
| GPT-4o-mini | ~10-15% | Internal testing |
These numbers are rough estimates that vary by domain. Medical and legal questions hallucinate more frequently than common knowledge questions. Obscure topics hallucinate more than popular ones.
The trend is improving. GPT-3.5 hallucinated 15-25% of the time on factual queries. Today’s best models are under 5% for straightforward questions. But even 3% means 1 in 33 factual statements might be wrong.
Types of Hallucination
Fabricated facts. The model states something untrue as fact. “The Eiffel Tower is 324 meters tall” (correct) vs. “The Burj Khalifa was completed in 2008” (it was 2010).
Fabricated sources. The model invents citations, URLs, or references that don’t exist. This is especially common when asked to provide academic sources.
Confident extrapolation. The model extends real patterns into false territory. If it knows a company’s revenue for 2023 and 2024, it might state 2025 revenue as fact rather than acknowledging it doesn’t know.
Entity confusion. The model mixes up details between similar entities. It might attribute a quote by one CEO to another, or merge biographies of two people with the same name.
ELI5: Next-Token Prediction — AI language models work by predicting one word at a time. They read everything written so far and guess the most likely next word. Then they add that word and guess the next one. And the next. It’s like an extremely sophisticated autocomplete on your phone — except instead of completing one word, it completes entire paragraphs. The “intelligence” emerges from doing this prediction really, really well.
How to Reduce Hallucination
1. Use RAG (Retrieval-Augmented Generation). Give the model your documents to reference before answering. Models hallucinate less when they have source material to quote. See our full RAG explainer.
2. Ask the model to cite sources. “Answer based on the provided documents. Quote the relevant sections. If the answer isn’t in the documents, say ‘I don’t have enough information to answer this.’”
3. Use system prompts to encourage honesty. “If you’re not confident about a fact, explicitly say so. It’s better to say ‘I’m not sure’ than to state something incorrect.”
4. Verify independently. For anything consequential — legal, medical, financial, academic — independently verify the model’s claims. AI is a first-draft tool, not an oracle.
5. Use the right model. Larger, more capable models hallucinate less. Claude Opus 4 and GPT-4o hallucinate significantly less than their smaller, cheaper counterparts. For high-stakes tasks, the cost premium of a better model is worth the accuracy improvement.
6. Lower the temperature. Temperature controls randomness in generation. Higher temperature (more creative) = more hallucination. Lower temperature (more deterministic) = less hallucination. For factual tasks, use temperature 0 or 0.1.
ELI5: Temperature — Temperature is a creativity dial for AI. Turn it down to 0 and the AI gives the most predictable, “safe” answer every time. Turn it up to 1 and the AI gets more creative and unpredictable — great for brainstorming, bad for facts. It’s like the difference between asking a careful accountant vs. an improvisational comedian to answer your question.
The Legal and Professional Implications
AI hallucination has already caused real-world problems:
- Lawyers have submitted court filings with AI-fabricated case citations
- News articles have been published with AI-generated statistics that don’t exist
- Medical chatbots have provided incorrect dosage information
- Customer support bots have made up company policies that don’t exist
The liability pattern is clear: if you deploy AI without verification guardrails, you own the consequences of its hallucinations. “The AI said it” is not a legal defense.
Will Hallucination Be Solved?
It’s improving but unlikely to be fully eliminated. The fundamental architecture of language models — next-token prediction based on statistical patterns — means there will always be cases where the model generates plausible but incorrect text.
The practical solution isn’t eliminating hallucination. It’s building systems that:
- Ground responses in source documents (RAG)
- Flag uncertainty rather than fabricating confidence
- Include human verification for high-stakes outputs
- Use multiple models to cross-check important claims
Current state: AI models are reliable enough for first drafts, brainstorming, coding assistance, and routine tasks. They’re not reliable enough for unsupervised factual claims in legal, medical, financial, or academic contexts.
For a detailed comparison of which models are most reliable, see our model leaderboard.