Why AI makes things up, and how to catch it

Sooner or later, every AI user meets a confident, well-written answer that turns out to be false. A quote from a book that does not contain it. A safety guideline that no organization ever issued. A tidy summary of your own document that includes a point the document never makes. People call this hallucination, and for a membership site it is the single biggest risk in using AI for content.

Your members pay for information they can trust. One invented fact in a lesson or newsletter can undo years of credibility, especially if members act on it. The good news is that invented answers follow patterns, and a few simple habits catch most of them before they reach anyone.

Why AI makes things up

An AI assistant does not look facts up in a database the way a search engine finds a page. It generates text by predicting, word by word, what a good answer would most likely say, based on patterns it learned from enormous amounts of writing. Most of the time, the most likely answer is also the correct one. But when the tool lacks the information, it does not naturally stop and say so. It produces the most plausible-sounding answer instead.

That is why hallucinations are so convincing. They are not random nonsense; they are exactly what a correct answer would look like. A made-up study has a believable title. An invented statistic is a sensible-looking number. The confidence of the writing tells you nothing about its accuracy.

The kinds of mistakes to expect

In our experience, invented content tends to fall into a few familiar types:

• Invented sources: books, articles, studies, court cases or web pages that do not exist, or real ones that say something different.
• Wrong specifics: dates, measurements, dosages, prices, names and numbers that are slightly or completely off.
• Outdated information: rules, product details or practices that have since changed.
• Blended facts: two real things merged into one false one, such as one organization's policy attributed to another.
• Misread documents: summaries of your own files that add, drop or twist a point.
• Filled gaps: details about your membership, such as a lesson title or an event date, simply invented because you did not supply them.

Warning signs in AI output

Hal runs a woodworking membership. He learned to slow down whenever a draft contained any of the following, and he now catches most problems on the first read:

• Very precise numbers with no source, such as exact drying times or load limits.
• Quotations, especially neat ones attributed to famous people.
• Named studies, reports, experts or organizations.
• Anything about safety, health, law, money or professional standards.
• Details about your own membership that you did not provide.
• An answer that fits your argument suspiciously perfectly.

None of these means the content is wrong. They mean it must be checked before you use it.

Habits that catch invented facts

Check claims against the original source

For any fact you plan to publish, find it in a source you trust: the manufacturer's instructions, the official body's own website, the book itself. If you cannot find it, cut it or rewrite the sentence so it does not depend on it.

Ask the AI to list its own claims

Before you verify, have the assistant pull out every checkable statement, so nothing slips past you:

Review the draft below. List every factual claim it makes, including numbers, names, dates, quotations, sources and instructions. For each one, say whether it came from the material I gave you or from your general knowledge, and rate how confident you are. Do not rewrite the draft. Draft: [paste the draft]. Material I provided: [paste your source notes]

A good result is a numbered list you can work through with a pen. Treat the confidence ratings as a rough guide only; the tool can be confidently wrong. What matters is the list, which you then check yourself.

Treat AI-supplied sources as leads, not proof

If an assistant gives you a reference, open it and confirm it exists and says what was claimed. Web-connected tools that show links make this easier, but even a real link can be summarized inaccurately.

Prevent errors at the start

You can reduce invention before it happens. Give the AI the material to work from rather than asking it to recall facts, a habit covered in briefing AI like a new assistant. Tell it plainly: “Use only the information provided. If something is missing, say so rather than guessing.” Ask it to mark any sentence it is unsure about. And keep AI for jobs where it shines, like structure and wording, while the facts come from you. The article on what AI can and cannot do maps out where those lines usually fall.

Clear instructions help too. Vague requests invite the tool to fill gaps, which is where invention creeps in; prompting basics shows how to close them.

Let the stakes decide how hard you check

Not everything needs the same scrutiny. A brainstorm of discussion topics needs a glance. A newsletter needs its facts checked. Content members may act on needs the strictest review of all.

Mei runs a continuing-education membership for dental hygienists. For her, any clinical statement is checked against the relevant professional guidance by a qualified reviewer before publication, whatever tool helped draft it. If your content touches health, law, finance or professional rules, keep it general and check what applies with a qualified professional. AI can help you explain; it should never be your authority.

Build the habit

• Add the warning-sign list to wherever you keep your prompts.
• Run the claim-listing prompt on your next AI-assisted draft and verify each item.
• Add “say so rather than guessing” to your standard instructions.
• Decide which of your content types need a second reviewer.
• Keep a short log of any errors you catch. Patterns will show you where to look hardest.

Checking takes minutes. Rebuilding members' trust after they act on a made-up fact takes far longer.