How to check an AI answer: five steps to verify what a chatbot tells you

A chatbot gives every answer in the same confident tone, whether it's right or made up. You don't need to check everything it says, but you do need a quick habit for the answers that matter. Here's what is going on behind an AI answer, a five-step check that takes a few minutes, and prompts that make it easier.

Why AI answers need checking

A language model writes by predicting what words should come next. That makes it fluent, but fluent isn’t the same as correct. When it doesn’t know something, it often fills the gap with something that sounds right: a hallucination. OpenAI’s researchers argued in September 2025 that one reason is how models are trained and tested: the scoring rewards a lucky guess more than an honest “I don’t know”.

Real tests show how often this happens. Both of the studies above looked at news, so the numbers are a warning sign, not an error rate for every kind of question:

In a study published in March 2025, Columbia University’s Tow Center gave eight AI search tools a quote from a news article and asked for its headline, publisher, date and link. Collectively, they were wrong on more than 60% of queries, from 37% for Perplexity to 94% for Grok 3, and they rarely admitted they weren’t sure.
In a study published in October 2025, journalists from 22 public broadcasters checked more than 3,000 answers about the news from ChatGPT, Copilot, Gemini and Perplexity. 45% had at least one significant issue, and 31% had serious sourcing problems: sources that were missing, misleading or wrong.

What the chatbot knows, and what it looked up

A chatbot can answer in two ways, and it helps to know which one you’re getting.

From memory. Everything the model learned in training, up to its knowledge cutoff. It can be out of date, and it can blend facts together.
From a search. Many chatbots can now search the web first and answer from what they find. This is called grounding. ChatGPT, for example, searches automatically when your question would benefit from current information, or you can switch search on yourself.

Links usually mean the chatbot searched, but models can also make up links, so a link only counts once you’ve opened it. And a search isn’t a guarantee: the model can still misread a page or mix up two sources. The Tow Center also found tools making up links and pointing to copies of articles instead of the originals.

The five-step check

Decide how much it matters. A dinner idea or a first draft doesn’t need checking. Health, money, legal questions, school or work facts, and anything you’ll share or act on do.
Open the sources. Ask for them if there are none. Then click: does the page exist, and does it actually say what the AI claims? Check that it’s the original, such as the study, the official page or the publisher’s own article, not a copy or a summary of it.
Find a second source. Look for the key fact somewhere independent of the chatbot: the organisation’s own website, a government page, or a trusted news outlet. If you can’t find it anywhere else, treat it as unconfirmed.
Check the date. Is the information current? Prices, rules, product features, job titles and “the latest” anything change often.
Be wary of neat details. Exact figures, quotes, names, book titles and legal cases are easy for chatbots to get wrong, often because such one-off facts appear rarely in the text they learned from, and a wrong one looks just as convincing as a right one. Lawyers have been fined for citing court cases a chatbot made up (see AI fails). Check each one.

Don’t rely on the chatbot’s own confidence. A model that says “I’m certain” is not necessarily more likely to be right: see calibration.

Prompts that make checking easier

These don’t make an answer correct, but they make it much quicker to check:

“Search the web and give me a link for each fact.”
“Answer only from the document I pasted. If it doesn’t say, tell me.”
“Which parts of your answer are you least sure about, and where should I check them?”
“Quote the exact sentence from the source that supports this.”

The second one helps most: when the AI works only from a document you trust, there is far less room for made-up facts. That’s the idea behind RAG.

When not to rely on AI

Use AI as a starting point, not the last word, for medical, legal and financial decisions: check with a doctor, lawyer or adviser before you act. Don’t upload private documents, such as medical reports or bank statements, without checking what happens to them first (AI privacy). And if a message, voice or video asks you to act fast, check it outside the chat (spot AI fakes and scams).

How we put this guide together

The figures come from the original studies, linked next to each one: the Tow Center for Digital Journalism at Columbia University (March 2025), the European Broadcasting Union and BBC study of AI assistants and the news (October 2025), and OpenAI’s research on why language models hallucinate (September 2025). The five steps are our own summary of common fact-checking practice. Got a checking habit that works for you? Tell us.