Menu

Why AI Makes Things Up (and How to Catch It)

Updated 2026-10-12

Ask an AI assistant for a statistic, a quote, or a list of sources and you’ll often get a clean, confident answer. Most of the time it’s useful. Sometimes it’s simply wrong, and it sounds exactly as confident as when it’s right. That’s the single most important thing to understand before you rely on AI for anything that matters.

This explainer covers what these errors are, why they happen, where they show up most, and the habits that catch them.

What a “hallucination” is

“Hallucination” is the common name for when an AI model produces output that is fluent and confident but false: a made-up statistic, a quote nobody said, a research paper that doesn’t exist, a product feature the product doesn’t have, or a link to a page that was never published.

The key word is confident. A person who doesn’t know something usually hesitates. AI models often don’t, unless you make it easy and expected for them to say “I don’t know.”

Why it happens

Large language models are trained to produce likely-sounding text based on patterns in enormous amounts of data. They are not looking facts up in a verified database unless they’re connected to a search or retrieval tool, and even then they can misread or misstate what they found.

OpenAI’s research write-up, Why language models hallucinate (and the accompanying paper), argues that a big part of the problem is incentives: standard training and evaluation reward models for guessing rather than acknowledging uncertainty. Their analogy is a multiple-choice test: a wild guess might score a point, while leaving the answer blank guarantees zero. Models graded mainly on accuracy learn to guess.

In everyday use, three practical causes stand out:

  1. Gaps in your prompt get filled with guesses. If you ask for an About page and don’t say when the business was founded, the model may invent a year.
  2. Specific details are the weakest point. Names, numbers, dates, prices, and citations are exactly where plausible-but-wrong output is most likely.
  3. Information goes stale. A model’s training data has a cutoff. Without web access, it can’t know about recent changes, and may not tell you it doesn’t know.

Where it bites hardest

AreaTypical errorWhy it’s costly
Statistics“73% of customers prefer…” with no real sourceMisleads readers; damages trust and credibility
QuotesWords attributed to a real person who never said themReputational and potentially legal risk
CitationsReal-sounding papers, authors, or URLs that don’t existEmbarrassing if published; wastes research time
Product features and pricesOutdated or invented plan detailsMisleads buyers; a problem on affiliate pages
Legal, medical, financial detailsWrong rules, dosages, deadlinesReal harm
CodeFunctions or library options that don’t existBroken builds, security issues

Prevention: build it into the prompt

You can’t eliminate hallucinations with prompting, but you can cut them down a lot.

Give the facts yourself. The best way to prevent invented details is to supply the real ones. Paste your notes, policies, product details, or source documents and say “use only these.”

Make “I don’t know” an acceptable answer. Add a line like:

Use only the information I've provided. If something is missing or you're unsure,
write [NEED INFO] instead of guessing. Do not invent statistics, quotes, or sources.

This is built into many of the templates in our Prompt Library for exactly this reason.

Ask for sources up front. “Cite a source for every factual claim, and prefer primary sources” makes claims easier to check. It doesn’t guarantee the sources are real, which is why you still open them.

Use web-connected tools for current topics. A tool that can search, such as Grok, is better for recent events than a model working only from memory. It can still misread a page, so verification still applies.

Detection: check after

A practical verification routine that takes a few minutes:

  1. List the claims. Ask the model: “List every factual claim in this text as a numbered list.” This turns a smooth paragraph into a checklist.
  2. Open every source. Does the link work? Does the page actually say that? Is it current?
  3. Cross-check numbers against a second independent source, especially anything you’ll publish.
  4. Ask for a self-critique. “Review your answer. Which claims are you least sure about, and why?” Models are often decent at flagging their own weak spots when asked directly, though this is a helper, not proof.
  5. Search quotes. Paste any quote into a search engine in quotation marks. If you can’t find it in a reliable source, don’t use it.
  6. Test code. Run it. Don’t trust code that has only been read.

Example

What this means for publishing

If you publish AI-assisted content (on a blog, in ads, in emails), you’re responsible for its accuracy. Google’s guidance on helpful, reliable, people-first content focuses on trustworthiness regardless of how content is produced. On eTarget, we don’t publish statistics, prices, or quotes we haven’t verified, and we mark unverified items for review instead of guessing.

Checklist before you trust or publish an AI answer