AI tools make things up because they generate text by predicting what words are likely to come next, not by looking up verified facts — so when the model has no reliable pattern to draw on, it produces a confident-sounding sentence anyway.
This is called a hallucination, and the single most useful habit you can build is to sort every answer into two piles: claims you can check in a few seconds, and claims that would cost you real money or reputation if they were wrong. The first pile you skim. The second pile you verify before you act.
What is actually happening inside the model
A language model is a statistical pattern machine. During training it reads enormous amounts of text and learns which words tend to follow which other words. When you ask a question, it does not query a database of truths.
It generates the most plausible continuation of your prompt, one chunk of text at a time. Most of the time that process lands on something accurate, because accurate text is common in the training data. But the model has no internal flag that says "I don't know this."
Plausibility and truth are two different things, and the model only optimises for the first one.
The gaps matter more than the model's overall quality. A model can be excellent at explaining photosynthesis and terrible at naming the current CEO of a small company, because one topic appears in millions of documents and the other appears in almost none. When the pattern is thin, the model fills the hole with something that sounds right. That is why hallucinations cluster around specifics: names, dates, version numbers, prices, citations, and statistics. General explanations are usually safe. Precise details are where the fabrications live.
A worked example you can recognise
Suppose you ask an AI assistant to recommend a project management tool for a five-person software team and to include current pricing. It might answer with a confident paragraph naming Linear and describing a "Starter" plan at a specific monthly rate per user. Linear is a real product — our AI tool database lists it as developed by Linear Inc., with a free tier, a Basic tier, and a Business tier, and an editorial rating of 4.7 out of 5.
But if the answer invents a plan name or a number that isn't on Linear's own pricing page, that detail is fabricated even though the surrounding recommendation is reasonable. The tool is real; the price is a guess wearing a suit.
The same trap appears with Notion AI. Our database records Notion AI as an add-on priced separately from the workspace plans, with the workspace itself offered on Free, Plus, and Business tiers. An AI answer that blends those two things together — quoting one number as if it covered everything — is wrong in a way that looks tidy. This is the pattern to watch for: the more specific and confident the number, the more likely it was generated to fill a gap rather than retrieved from a source.
A decision rule for what to verify first
Not every claim deserves a fact-check. Use this order. First, verify anything that would change a decision: prices, legal or medical specifics, API names, version numbers, and anything you plan to paste into a document with your name on it. Second, verify anything with a number attached, because numbers are the cheapest thing for a model to invent and the easiest for you to check. Third, verify anything attributed to a person or publication, since fabricated quotes and citations are common.
Skip verification for brainstorming, rewrites, tone adjustments, and explanations of well-known concepts. If the model explains what a for-loop does, you do not need a second opinion.
One more habit helps: ask for the same specific fact twice in two separate conversations. A model that knows something tends to repeat it consistently. A model that is guessing tends to produce a slightly different guess the second time. That inconsistency is your signal.
Where this advice breaks down
Verification costs time, and for high-volume work the cost adds up fast. If you are generating fifty product descriptions, checking every claim is not realistic — so restrict AI to the parts that don't carry factual risk, like tone and structure, and write the facts yourself. Also, a confident tone is not evidence of accuracy, and neither is a long answer.
Length and certainty are style, not proof. Finally, some tools now cite their sources; treat those citations as a starting point, not a guarantee, because a citation can itself be fabricated or point to a page that says something different from what the answer claims. The honest summary is this: AI is a fast first drafter, not a source of record. You are still the one signing off.