What you’ll be able to do

By the end of this you’ll know why AI assistants invent facts, and you’ll have five habits that catch it before it costs you anything. Not zero hallucination, mind. Nobody can promise you that, and anyone who does is lying to you already. Just fewer surprises.

Before you start

Here’s the thing nobody tells beginners: when an AI assistant doesn’t know something, it doesn’t say “I don’t know.” It writes something that sounds right. That’s hallucination, and it’s not a bug that got missed, it’s just how these models work.

ChatGPT, Claude, Gemini, Copilot, all of them, generate text by predicting what word comes next based on patterns. Most of the time that produces something true, because most patterns in the training data were true. But when the model hits a gap, a citation it never actually saw, a court case, a statistic, a quote, it doesn’t stop. It fills the gap with something plausible. It’s not lying, exactly. It doesn’t know the difference between a real quote and a convincing one.

This shows up worst with specifics. Names, numbers, quotes, case law, product features, that sort of thing. Ask a model to summarise a well known idea and it’ll usually do fine. Ask it for the exact statistic from a study, or the precise wording of a court ruling, and you’re in riskier territory.

And this happens in every mainstream assistant right now. Not one of them has solved it. Some are better at flagging uncertainty than others, but none of them are immune. Worth sitting with that before you start.

Get set up

You don’t need anything complicated to start. A free account with one of the big assistants will do; ChatGPT, Claude, Gemini or Copilot all work for what follows. But which one you pick, and which mode you use inside it, matters more than you’d think.

If reducing made-up facts is your main goal, I’d point you towards tools that answer from actual retrieved sources rather than from memory alone. Perplexity is built around this: it searches the web and shows you what it found, with links, rather than just generating an answer and hoping. NotebookLM works differently but for the same reason, it answers only from documents you upload, so it can’t wander off into invented territory the way a general chatbot can.

That’s not to say ChatGPT, Claude, and Gemini are useless here. Each of them now has some form of web search built in (ChatGPT and Copilot search the web, Gemini does too, Claude has search available in some plans). Turning that on, when it’s available, changes the game, because the model is grounding its answer in something it can point to, rather than just recalling.

So: for quick chat and drafting, any of the big four is fine. For anything where the facts genuinely matter, lean on search mode, or use Perplexity or NotebookLM specifically. Know what you’re using and why before you start typing.

Try it yourself

Right, let’s actually see this happen. Here are three prompts. Run them as they are, in whatever assistant you’ve got open.

First, a prompt designed to expose the gap. Ask for something specific that the model might not actually know:

Give me a direct quote from [a specific person] about [a specific topic], including the exact source and date.

Watch what comes back. If it gives you a confident quote with a tidy citation, be suspicious. That confidence is not the same as accuracy.

Second, a prompt that asks the model to admit uncertainty instead of guessing:

Answer the following question. If you are not fully certain of any part of your answer, say so explicitly and explain what you're unsure about, rather than guessing: [your question]

This one’s doing real work. You’re not asking for a better answer, you’re asking for an honest one. Some models respond well to this. Some still slip back into confident-sounding filler. That’s useful information either way.

Third, a prompt that forces sourcing:

Answer this question: [your question]. For every factual claim in your answer, tell me where that information came from, and rate your confidence in it as high, medium, or low.

That last one is the habit I’d keep for life. Confidence per claim. It doesn’t fix the underlying problem, but it gives you something to check against, and it makes the model slow down a bit rather than just riffing.

Try all three on the same question and compare. You’ll notice they pull different behaviour out of the same tool.

Check the result

Right, so you’ve got an answer. Now what.

First test: click the citations. If the assistant gave you a link, open it. Does the page actually say what the assistant claims it says? Summaries routinely attribute a stat to a source that, when you actually read it, says nothing of the sort.

Second test: ask for sources explicitly if it didn’t give you any.

List the specific sources you used to answer that, with links where possible.

If it can’t produce real ones, or the links don’t resolve, or they resolve to something unrelated, that’s your answer. The original claim was probably invented.

Third test, and this one’s underrated: ask the same question twice, in separate conversations, and compare the answers. If the facts shift between runs, particularly names, dates, or numbers, that’s a strong signal you’re dealing with generated plausibility rather than retrieved fact. A real fact tends to stay the same. A hallucinated one wanders.

Fourth: read the answer specifically looking for anything with a number, a name, or a quote in it. Those are the danger zones. Everything else in the answer might be fine, general explanation, reasoning, structure. But numbers, names, quotes: check those, always, before you repeat them to anyone else.

None of this takes long once it’s a habit. A few minutes of checking beats an embarrassing correction later.

If it doesn’t work

Sometimes you do all this and the model still gets it wrong confidently. That’s not you doing it badly, that’s the current state of the technology. No setting eliminates hallucination. Not search mode, not confidence-per-claim prompts, not a paid subscription. They reduce it. They don’t remove it.

Here’s what matters more, though: what you don’t hand over in the first place.

Don’t paste medical records into a general assistant expecting private, careful handling of sensitive detail, and don’t treat its medical answer as diagnosis. Don’t upload legal documents with client details, contract terms, or anything under NDA, unless you actually know how that tool handles data retention, because “I asked ChatGPT to review my contract” can go wrong in two separate ways: the advice might be invented, and the document itself might now sit somewhere you didn’t intend. Don’t feed in proprietary business data, internal financials, unreleased product details, anything your employer would wince at seeing in someone else’s training set, unless you’re on an enterprise plan with data protections you’ve actually checked yourself, not assumed.

Read the actual privacy policy of whatever tool you’re using before you decide what’s safe to type into it. Not the marketing page. The privacy policy.

And go back to web search settings, because this trips people up constantly. If your assistant has search available but it’s switched off, or you’re on a mode that doesn’t use it, you’re getting memory-based answers dressed up to look current. Check the toggle. It’s usually right there, and it’s easy to miss.

Keep exploring

If this has been useful, there’s more worth reading. “How to get AI to be honest with me” goes deeper into the confidence and hedging side of things. “How to get AI to fact-check something” walks through checking a specific claim rather than a whole answer. And “How to get AI to help me research a topic” covers building up a proper research habit rather than one-off questions.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.