Why AI makes things up (hallucinations) and how to reduce it

What's actually happening

A language model predicts text that fits the pattern of your request. When the right answer is well represented in its training or your input, the prediction is often correct. When it isn't, the model can still produce fluent text — it has no built-in signal that says “I don't know.”

When hallucinations are most likely

Watch for these high-risk requests.

1. Specific facts not in your input: dates, figures, names, quotes.
2. Citations, links, and references to papers or laws.
3. Recent events beyond what the tool can access.
4. Niche or local details, such as a small company's policy.
5. Long outputs where small errors hide in plausible text.

Techniques that lower the risk

Give the model the source and restrict it to that source. Ask it to quote the sentence each claim relies on. Tell it what to do when information is missing. Break long tasks into smaller checked steps. None of these guarantees accuracy, but together they make errors easier to catch.

Use only the policy text below. For each answer, quote the sentence you relied on. If the policy does not cover the question, reply “not covered by this policy”.

Always verify what matters

For anything that affects money, health, legal standing, or someone's reputation, verify against a primary source. If you can't verify a claim, leave it out or label it clearly as uncertain.

Mistakes that make hallucinations more likely

Some everyday habits invite invented answers.

1. Asking leading questions (“Why did sales fall in March?”) when you haven't confirmed that they fell.
2. Requesting citations without supplying sources, then trusting the ones produced.
3. Asking for very long outputs in one go, where errors hide in the middle.
4. Treating a follow-up “Are you sure?” as verification — the model may simply change its answer.

A quick verification routine you can use daily

Before using an AI answer, spend two minutes on three checks. First, underline every specific claim: names, numbers, dates, quotes. Second, for each underlined claim, find where it came from — your source, or nowhere. Third, for claims from nowhere that matter, check a primary source or remove them.

This routine becomes fast with practice and catches most serious errors. For high-stakes work, add a second person's review. The goal isn't to distrust everything, but to know exactly which parts of an answer you've confirmed.