LIVE
Court filings reveal internal admissions on AI content scraping19/09/26|US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Court filings reveal internal admissions on AI content scraping19/09/26|US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|
Intermediate🌀

Understanding AI hallucinations

Why does ChatGPT invent references? Why did Bard make Google lose $100 billion? We explain how AI hallucinations work, how to detect and avoid them.

11 min readPublished May 5, 2026· Updated September 17, 2026

In one sentence

AI hallucinations are when an LLM invents false but plausible information and delivers it to you confidently as if it were true. Understanding why it happens and how to detect them has become essential.

🌀
The analogy that works
Imagine a brilliant student at an oral exam who hasn't revised. They know they need to answer quickly, their reputation is on the line. So they bluff with confidence: they make up a date, attribute a quote to the wrong person, mix up two concepts. Everything sounds right, except it's wrong. That's exactly how an AI hallucinates.

🔍 Want an AI that cites its sources?

Compare models with integrated web search: Perplexity, Claude with MCP, Gemini grounded.

See the comparison

Why do LLMs hallucinate?

To understand, you need to remember how it works: an LLM predicts the next word based on the billions of sentences it's read. It doesn't have a verified knowledge base, it just has statistical probabilities.

The exact mechanism
When you ask it "Cite me a study on X", it looks at what resembles study citations in its data. It then assembles a credible title, a plausible author name, a known scientific journal, a coherent date. The result: a citation that doesn't exist but sounds 100% true. The model doesn't know it's inventing. It's just doing its job: producing plausible text.

The 4 types of hallucinations

Not all hallucinations are equal. Here's the typology:

The 4 types of hallucinations

 🌀Type📝Typical example
FactualFalse facts, dates, figures'Einstein was born in 1880' (false: 1879)
Fabricated sourceCitations of non-existent works'According to Smith et al. 2019, Journal of AI...' (doesn't exist)
LogicalFalse but convincing reasoning'If A then B, therefore C' with C not following
ContextualConfusion between provided informationYou provide 5 docs, it invents a detail between the 2

How often does it happen?

It's more frequent than you think. The official figures:

Hallucination rate by model (2026 benchmarks)

GPT-3.527%
GPT-4o12%
Mistral Large 38%
Gemini 3 Pro6%
GPT-54%
Claude Opus 4.73%

Reading: even the best 2026 models hallucinate 3-4% of the time on factual questions. Out of 100 answers, 3-4 contain an error. Out of 1,000 answers: 30-40 errors. That's not negligible.

When does it happen most?

Hallucination red zones
1. Niche / specialized domains Rare medicine, local laws, sector statistics. The model has seen little, it cobbles together. 2. Recent facts (post-cutoff) Its knowledge stops at its training date. For recent news, it invents or refuses. 3. Very precise questions with numbers "How many inhabitants in Vesoul in 2024?" → high risk of hallucination if the exact figure isn't widespread on the web. 4. Citations and sources "Give me 3 scientific references on X" → maximum danger. The model often invents papers that don't exist. 5. Obscure or poorly documented code For mainstream languages (Python, JS), good. For ultra-specific libs, it invents functions that don't exist.

Hallucinations that caused scandals

📚3 real cases that are scary

1. The lawyer who cited 6 invented case laws (2023)

A New York lawyer used ChatGPT to prepare a case file. He submitted to the judge 6 case laws that ChatGPT had... completely invented. The judge verified, sanctions fell: fine, professional reprimand.

Lesson: NEVER submit official AI content without verification.

2. Air Canada ordered to honour its chatbot's promises (2024)

Air Canada's chatbot invented a refund policy that didn't exist. A customer relied on it to buy a ticket. Air Canada refused to honour it. The court ruled: the company must refund, the chatbot is their responsibility.

Lesson: a company chatbot is legal liability for the company.

3. Google Bard at 100 billion (2023)

During its launch demo, Bard (Google) claimed that the James Webb telescope had taken the first photo of an exoplanet. That's false (it was another telescope, in 2004). The error was spotted by astronomers in real time. Google stock -8%, or 100 billion in market cap lost.

Lesson: hallucinations cost really expensive in the real world.

How to detect a hallucination

Not easy, by design (it's plausible). But here are the red flags:

Anti-hallucination checklist
🔍 Ask for sources "Where do these figures come from?" If the AI answers vaguely ("from recent studies"), it's suspect. If it cites a specific source, verify that the source actually exists. 🔍 Cross-check with Google Important facts → Google. If the figures don't match, doubt. 🔍 Beware of excessive precision "Karpathy's 2019 study shows that 73.4% of LLMs..." → too precise, often invented. Real studies are rarely that round. 🔍 Ask 2 different models ChatGPT + Claude. If both say the same thing, more chance of being true. If divergence, doubt. 🔍 For code: test it If the LLM uses an obscure function, run the code. Many errors: "TypeError: function does not exist".

Technical solutions to reduce hallucinations

Technical solutions 2026

 🛠️Solution📉Hallucination reduction
RAG (Retrieval Augmented Generation)Forces the model to cite a database-70% to -90%
Chain-of-thought promptingForce step-by-step reasoning-30% to -50%
Self-consistencyGenerate 5 answers, take the majority-20% to -40%
Integrated web searchReal-time connection to Google/Bing-60% to -80%
Reasoning models (o1, Opus thinking)The model 'thinks' before responding-40% to -60%
Fine-tuning on your dataSpecialize the model on your domain-50% on your domain

RAG: the star solution

RAG (Retrieval Augmented Generation) has become the technique to limit hallucinations in enterprises. The principle:

How RAG works

  1. You ask a question

    The system receives your prompt.

  2. Search in a database

    The system searches your documents (PDFs, database) for relevant passages.

  3. Context injection

    The found passages are sent to the LLM along with the question.

  4. Anchored response

    The LLM responds based on the provided documents, with citations.

Result: instead of inventing, the LLM cites real passages. Hallucinations divided by 5 to 10.

📚 Want to understand RAG in detail?

We explain RAG simply, with the tools and costs to implement it.

Read the RAG article

Use cases where NOT to trust an LLM alone

Red zone : always verify elsewhere
- Medical: symptoms, diagnoses, dosage → always doctor - Legal: laws, case laws, contracts → always lawyer - Financial: investment advice, tax calculations → always advisor - Security: protection advice, vulnerabilities → always security expert - Academic research: citations, references → always verify sources - Critical decisions: hiring, firing, sanctions → always human For all of these, AI can help you structure, brainstorm, summarize. But NEVER the final decision.

The metaphor that sums it all up

🎤
The expert on hot mic
Imagine an expert guest on live radio being questioned on 100 different subjects. They know 80% of the subjects well. But on the remaining 20%, they can't say 'I don't know' (forbidden by their company). So they bluff intelligently: they combine their knowledge to produce answers that sound credible. On the 20%, they have 40-60% correct answers by chance : and 40-60% of falsehoods presented as truth. That's exactly an LLM. It doesn't bluff out of malice, it bluffs because it doesn't have the right (statistically) to just stay silent.

Absolutely remember

  • All LLMs hallucinate: 3-4% for the best, 10-25% for older ones
  • ✅ Risk zones: niche, recent, citations, precise figures, obscure code
  • Always verify critical facts, especially for professional use
  • ✅ To reduce: RAG, chain-of-thought, reasoning models, web search
  • Never a final decision on medical, legal, financial based on AI alone

Understanding hallucinations means avoiding disasters. AI is a great assistant, not an oracle.

🧠 Quiz
Question 1 of 3

Why does an LLM hallucinate?

To go further

Tags
HallucinationsFiabilitéRAGLLM

Read next