AI confidently told someone that the Taj Mahal was built in 1632 — wrong year. It cited a research paper that does not exist. Here is why AI hallucinates, the 5 types of hallucination, and how to catch them before you share them.
A lawyer submitted a court filing written by ChatGPT. It cited six previous cases — all of them fabricated. The AI had invented case names, citations, and rulings that sounded perfectly real. The lawyer did not check. The judge was not amused. This is not a hypothetical — it happened in New York in 2023, and it is the most famous example of AI hallucination. But smaller versions of this happen millions of times every day. Here is why AI makes things up, and how to catch it.
What hallucination actually is
In AI terms, hallucination means: the model generates content that is factually incorrect, nonsensical, or unfaithful to the provided input, but presents it with the same confidence as correct information.
The AI is not lying. Lying requires intent — knowing the truth and choosing to say otherwise. The AI has no concept of truth. It is predicting the most statistically likely next word, given its training data and your prompt. Sometimes the most statistically likely word is factually wrong — and the AI says it anyway, with complete confidence.
The 5 types of AI hallucination
### 1. Factual hallucination
The AI states a 'fact' that is incorrect. 'The Taj Mahal was completed in 1632.' It was 1653. 'India's population is 1.2 billion.' It crossed 1.4 billion years ago. The numbers are close — they sound right — but they are wrong.
How to catch it: Any factual claim — a date, a number, a statistic, a name — should be cross-checked. If the AI does not provide a source, assume it may be wrong until verified.
### 2. Source fabrication
The AI cites a book, paper, or study that does not exist. 'According to a 2022 McKinsey report on AI productivity...' The report sounds real, the year is plausible, McKinsey does publish such things. But this specific report was invented by the AI.
How to catch it: If an AI cites a source, search for that exact source. If the first page of Google results does not find it, it probably does not exist. AI-fabricated sources often have plausible-sounding titles that return zero real search results.
### 3. Arithmetic and logic errors
AI can explain complex mathematical concepts beautifully — and then get the actual calculation wrong. It can produce a logically structured argument where the conclusion does not follow from the premises. The words are right; the reasoning is broken.
How to catch it: Run the numbers yourself. Follow the logic step by step. If the AI says 'This proves X,' ask: does the evidence it presented actually support X, or is the conclusion just statistically associated with those kinds of words?
### 4. Over-generalisation
The AI takes a specific fact and applies it broadly. 'The Indian government's 2026 Budget increased infrastructure spending.' True for certain ministries, not for others. The AI flattens nuance into a generalisation that is partially wrong.
How to catch it: Look for qualifying words. If the AI says 'always', 'never', 'all', or 'every', it is probably over-generalising. The real world has exceptions. AI tends to smooth them out.
### 5. Temporal hallucination
The AI gives information that was true at some point but is no longer current. 'The PPF interest rate is 7.1%.' True for Q1 FY 2026-27. But if you are reading this beyond that quarter, the rate may have changed. The AI does not know what it does not know.
How to catch it: Any time-sensitive claim — interest rates, tax slabs, deadlines, prices, product availability — should be verified against a current source. Add 'as of [date]' to your prompt to force the AI to qualify temporal claims.
Why hallucination happens
Hallucination is not a bug — it is a property of how these models work. They are trained to predict the next word, not to verify facts. Three specific reasons:
1. Training data contains errors. The internet is full of wrong information. The model learned from it. 2. The model has no fact-checking mechanism. When you ask a human a question, they can pause, think, and say 'I am not sure.' The AI cannot. It must generate the next word, and it will always generate something — even if that something is wrong. 3. Confidence and correctness are uncorrelated. The AI can be 99% confident and 0% correct. The fluency of its output says nothing about its accuracy.
How to reduce hallucinations in your own use
1. Ask for sources. Add to your prompt: 'Cite your sources. If you are unsure about a fact, say so rather than guessing.' 2. Constrain the AI. 'Answer based only on the information I provided below. Do not add external knowledge.' For research tasks, upload the document and ask the AI to work from that document only. 3. Use AI for what it is good at. Drafting, brainstorming, summarising, explaining — tasks where you review and edit the output. Do NOT use AI as a final authority for facts, especially in finance, law, or medicine. 4. The two-AI check. Ask the same factual question to two different AI models (ChatGPT and Claude, for example). If they agree, the answer is more likely (but not guaranteed) to be correct. If they disagree, one or both are wrong.
Frequently Asked Questions
### Do all AI models hallucinate? Yes. Every large language model hallucinates — ChatGPT, Claude, Gemini, DeepSeek. The rate varies by model and by topic, but no model is hallucination-free. The best models hallucinate less, not never.
### Can hallucinations be fixed? Partially. Techniques like RAG (Retrieval-Augmented Generation) — where the AI searches a trusted database before answering — reduce hallucination significantly. But fundamentally, probabilistic text generation will always have some error rate. Zero hallucination would require a fundamentally different approach to AI.
### How often does ChatGPT hallucinate? Estimates vary widely — from 3% to 27% depending on the topic and how hallucination is measured. On factual questions in domains well-covered in training data, hallucination rates are low (under 5%). On niche, technical, or very recent topics, hallucination rates are significantly higher.
### Is AI hallucination dangerous? It can be. If you make a financial decision, a legal filing, or a medical choice based on hallucinated information, the consequences are real. The danger is not the hallucination itself — it is the human tendency to trust fluent, confident-sounding output.
### How do I know if an AI output is hallucinated? The only reliable method: verify against a trusted primary source. If the AI says 'the SCSS interest rate is 8.2%,' check the India Post website. If it cites a study, search for that study. If it gives a statistic, find the original data. There is no shortcut.
Disclaimer
This article is for educational purposes. AI output should always be verified before use in any consequential context.