Understanding AI Hallucinations (and How to Catch Them)

A lawyer in New York filed a court brief in 2023 citing six precedent cases. None of them existed. ChatGPT had invented all six, complete with fake judge names and fabricated quotes. The case got thrown out, the lawyer got fined, and the story went viral.

Three years later, professionals are still getting caught by the same trap. Doctors asking about drug interactions. Journalists citing fake studies. Students submitting essays with invented sources. The tools got better. The hallucinations did not go away. They got more plausible.

Here is what you need to know about why AI tools make things up, how often, and how to catch the lies before they catch you.

What a Hallucination Actually Is

An AI hallucination is when a language model generates text that sounds confident and plausible but is factually wrong. The model is not lying in any meaningful sense. It does not know things the way you or I do. It is predicting the next most likely word based on patterns in its training data.

When those patterns happen to produce a true statement, great. When they produce a confident-sounding false statement, that is a hallucination. The model has no internal alarm that distinguishes the two.

This is important because it explains why hallucinations are hardest to catch in cases where the model sounds most sure. A hedging, uncertain response ("I am not sure, but I think...") is less likely to mislead you than a clean declarative sentence delivered with full confidence. The confidence is the bug.

How Often It Happens (Honest Numbers)

Research from 2024-2025 measured hallucination rates across the big models on factual Q&A:

• GPT-4o: roughly 1-3% on common knowledge, up to 15-25% on niche or recent topics.
• Claude 3.5: roughly 1-2% on common knowledge, 10-20% on niche.
• Gemini 1.5: roughly 2-4% on common, similar on niche.
• Smaller open models (7B-13B): 5-15% on common, up to 40% on niche.

These numbers depend heavily on what you are asking about. For "what year was the French Revolution", hallucination is essentially zero. For "what did so-and-so say in a 2019 podcast episode about X", the numbers shoot up fast because there is less and less training signal.

Citations are particularly bad. Every major model will invent academic papers, URLs, and book titles. Model-generated citations should be treated as fiction by default until verified. This has not improved meaningfully despite three years of trying.

The Types of Hallucinations You Will Actually Hit

Fabricated citations. The classic. The model produces a citation, complete with authors, year, journal, and DOI. None of it is real. Any time a model gives you a source link, click it.

Outdated information presented as current. Models have a training cutoff. Ask one about a product that changed pricing last month and you will get confident wrong information. This is why Perplexity and other RAG-based tools are usually safer for anything time-sensitive: they pull from the live web.

Plausible but wrong code. This is the one that burns developers. The model will invent a function that does not exist in the library, or remember a deprecated API signature as current. Code compiles, runs, fails silently, and you spend an hour figuring out why.

Wrong numbers. Especially math that looks close to right. "The company reported $4.2 billion in 2023" when the real number was $3.8 billion. Small numeric errors sail through because they feel plausible.

Made-up historical details. Names, dates, relationships. A model might correctly tell you Napoleon was born in 1769 and then confidently tell you he went to an academy that did not exist yet.

Non-existent tools or features. Ask about a software feature that sounds like something the tool should have. The model may confidently describe the feature and how to use it. The feature does not exist.

Aggregation errors. Summarizing ten sources into five and getting three of them wrong. The summary reads fine until someone who knows the source material reads it.

Why the Problem Is Structural, Not a Bug

This is worth internalizing because it changes how you use these tools.

Language models are trained to produce fluent, coherent text. They are not trained to know what is true. The training process rewards plausibility above all else. When the model is uncertain, it has no way to express "I do not know" unless that phrase is statistically likely to come next in its output.

Researchers have been working on this for years. Techniques like retrieval-augmented generation (RAG), self-consistency checking, chain-of-thought prompting, and tool use all help. None of them eliminate the underlying issue. Models will keep inventing things because invention is a side effect of how they work.

This is why the best AI tools in 2026 are designed as co-pilots rather than oracles. Cursor, Claude, and others are most useful when you treat them as sharp interns with a vivid imagination, not as sources of ground truth.

How to Catch Hallucinations

Here are the habits that will save you. Build them into muscle memory.

Verify every citation before you use it. Every single one. Click the link. Search the paper on Google Scholar. If it does not exist, it does not exist.

Ask the model to cite sources, then verify them separately. Do not trust inline citations in chat. Use a search tool to confirm. Perplexity shows real citations by construction because it is searching the live web, but even there, read the source before repeating the claim.

For code, run it. Do not just glance at AI-generated code and ship it. Execute it. If it is a one-liner, paste into a REPL. If it is a function, write a quick test. Most code hallucinations fail fast when you actually run them, but you have to run them.

For numbers, reconcile with a primary source. Company financials? Go to their investor relations page. Population figures? UN or Census data. Historical dates? Wikipedia will do for a first pass; a real reference for anything high-stakes.

Ask the same question two different ways. If a model gives you a confident answer, rephrase the question and ask again in a new chat. Consistent answers across phrasings are more likely to be right. Inconsistent answers are a red flag.

Cross-check between models. Ask ChatGPT, Claude, and Gemini the same factual question. If all three agree, you are probably fine. If two agree and one disagrees, investigate. If all three give different answers, you are in hallucination territory and should not trust any of them.

Use grounded tools for factual work. NotebookLM only answers from documents you uploaded. Perplexity cites live sources. These tools are harder to hallucinate with because they are constrained to real material.

Watch for specificity mismatches. Real information gets fuzzier the more specific you push. "What did X say in their 2022 keynote?" should produce uncertain or hedged answers unless there is a widely indexed transcript. If the model gives you four exact quotes, be deeply suspicious.

Prompt Patterns That Reduce Hallucinations

You cannot eliminate hallucinations with prompting. You can reduce them meaningfully. These patterns actually work:

Give the model an out. End your prompt with "If you do not know or are uncertain, say so. Do not guess." This shifts the probability of hedged responses upward. It does not make the model honest, but it makes honesty more likely to win out.

Ask for citations upfront. "Provide a citation for each factual claim. If you cannot provide a real verifiable citation, leave the claim out." Models will sometimes drop claims rather than invent sources, which is exactly what you want.

Constrain the output to a source. "Answer using only the following text: [paste]. If the answer is not in the text, say 'not in source'." This is the prompting equivalent of RAG. It works.

Use step-by-step reasoning. "Think through this carefully, step by step, before giving your final answer." Chain-of-thought prompting reduces certain kinds of hallucinations, especially in math and logic.

Ask it to rate its own confidence. "Rate your confidence in each part of the answer from 1 to 10." Models will sometimes accurately flag low-confidence sections. Not always. But often enough to be useful.

Our guide on how to write better AI prompts goes deeper on this, including patterns for specific tasks.

When to Just Not Use AI for the Task

There are genuinely tasks where the hallucination risk outweighs the productivity gain. Know which ones they are for you.

• Medical diagnoses or drug information. Real consequences, and the training data is full of outdated edge cases. Use a proper medical reference.
• Legal citations, court cases, or statute references. Ask the lawyer from the intro.
• Investment or financial advice. Especially around specific stocks, tax implications, or regulatory rules.
• Anything where the downside of being wrong is much larger than the upside of being right fast. This is the general rule.

For everything else, AI is fine. It is a very smart intern. You would not let an intern send a client an unreviewed brief either.

The Actually Useful Frame

Hallucination is the tax you pay for fluency. The thing that makes language models feel like magic, the confident conversational voice, is the same thing that produces the confident fake citation. You do not get one without the other with current tech.

So use the tool. Get real leverage from it. But never skip the verification step when accuracy matters. Treat the model's output as a very good first draft that probably contains mistakes, not as a source of truth.

The people who get the most out of AI tools in 2026 are not the ones who trust them the most. They are the ones who have built efficient verification habits and can tell within seconds whether a response needs a fact-check.

Want to find tools that are less prone to hallucinations by design? Browse the research and analysis category for tools that ground their outputs in real sources, or look at the academic research helper stack for a full verification-first workflow.