Can you trust AI answers? How to fact-check what an AI tells you

You should not trust AI answers blindly: an assistant can state a mistake with exactly the same confidence as a fact. But a few simple habits, like asking for sources, opening them and going back to the official website, are enough in most cases to tell what is reliable from what is doubtful.

Why AI can be confidently wrong

Assistants such as ChatGPT, Gemini or Claude run on a large language model (LLM). The model does not look facts up in a database: it generates, word by word, the most likely continuation of the text based on what it has learned. Most of the time, the likely answer is also the right one. Sometimes it is not.

When AI makes up a number, a date, a quote or a reference, this is called a hallucination. The trap is that the tone never changes: an invented answer reads just as smoothly and confidently as an accurate one. Our article on why AI makes things up explains the mechanism in more detail.

Two other causes of error are common:

• The training cutoff. A model learns from texts collected up to a certain point, known as its knowledge cutoff date. Without web access, it does not know what has happened since and may present outdated information as current.
• A misread source. With AI web search, the assistant reads web pages before answering. That helps, but it can still pick an unreliable page, summarize it poorly or mix up two pieces of information.

When to check: it depends on the stakes

You do not need to check everything. A rewrite of your own text can be verified at a glance. A benefit amount, a medication dose or a legal deadline cannot. This table suggests a level of caution for each type of answer.

Type of answer	Caution	What to check	Where to check
Rewrite or proofreading of your own text	Low	Is the meaning unchanged?	Your original text
Explanation of a general concept	Medium	Does it match a reputable source?	Textbook, encyclopedia, institutional website
Numbers, dates, prices	High	Does the source exist and say the same thing?	Official body, public statistics
Health, law, money	High	What applies to your situation	Official website, professional
Recent news	High	The date and origin of the information	Official statement, original news outlet
Quote, book, scientific paper	High	Whether the reference exists at all	Library catalog, publisher’s website

Red flags that should make you check

Some signs should trigger an immediate check, even for a harmless question:

• Suspicious precision. A figure down to the cent, an exact date or a quote in quotation marks with no source at all.
• A reference you cannot find. A book, a statute, a study or a link that does not show up anywhere else.
• An answer that shifts. If you rephrase the question and a key point changes, neither version is safe.
• An answer that suits you a little too well. AI tends to go along with the way a question is framed. Ask it for the counterarguments too.
• A very recent or very local topic. A law passed last month, the opening hours of a local office, the rules of a sports club: this is where errors are most common.

The 5-step method

1. Ask for sources. Add to your question: “Cite your sources with a link for every important claim, and tell me when you are not sure.” An assistant without web search may not be able to do this, which is useful information in itself.

2. Open the sources. Click at least one link. Check that the page exists, that it is reputable and that it actually says what the assistant claims. A reference you cannot find is a red flag.

3. Go back to the primary source. For a government rule, go to the agency’s own website, such as gov.uk or the relevant .gov site. For a statistic, go to the organization that produces it. For a study, go to the publication itself. A blog post repeating a number is not a primary source.

4. Cross-check with an independent source. A second reliable source confirming the information increases your confidence. Note that asking another assistant is not an independent check, because it can repeat the same mistake.

5. Check the date. Information that was correct two years ago may be wrong today: prices, laws, opening hours, software versions. Look at the date on the page and prefer the most recent one.

Check yourself An assistant gives you a precise figure and a link to a news article. Is that enough?

Not quite. Open the link to confirm the article exists and actually cites that figure. Then trace the figure back to its origin: the organization or study the article mentions. Finally, check the date, because the figure may have been updated since.

Assistants that search the web

Assistants with web search usually cite the pages they used, which makes checking much faster. Here are the features of the main ones, as listed on the providers’ websites:

Assistant	Free	Paid from	Web search	Agent mode	Voice	Images
Perplexity	Limited	$10/mo	Yes	Yes	—	Yes
ChatGPT	Limited	$8/mo	Yes	Yes	Yes	Yes
Gemini	Limited	€4.99/mo	Yes	—	Yes	Yes
Claude	Limited	$20/mo	Yes	Yes	—	—
Microsoft Copilot	Limited	$9.99/mo	Yes	—	Yes	Yes

An empty cell means the provider does not specify the feature, not that it does not exist. Perplexity is designed first and foremost as a search engine with sources; according to the company, as of October 22, 2026 its free version offers nearly unlimited basic searches and 3 “Pro” searches per day. To compare other criteria, use our AI assistant comparison tool.

Prompts that make checking easier

How you phrase your question affects how easy the answer is to verify. A few useful prompts:

• “Separate what you are sure of from what you are assuming.”
• “Which parts of your answer should I check first?”
• “If you don’t know the answer, say so instead of guessing.”
• “Give me the exact name of the official document that contains this rule.”

These prompts do not eliminate errors, but they make them easier to spot. Also keep in mind that a web page read by an AI can contain hidden instructions meant to manipulate it: this is called prompt injection.

Images, video and voice: a special case

Text can be checked against a source; an image or a recording is trickier. A deepfake is an image, video or voice generated by AI, sometimes extremely realistic. When content is shocking or too good to be true, look for who published it first, run a reverse image search and wait for confirmation from a recognized source. Scammers exploit exactly this kind of doubt: our guide to AI scams shows how to spot them.

Building these habits as a family

These habits can be learned early. A child who asks AI for help with a school project can get used to opening one source and comparing it with their textbook. Our guide to kids and AI offers ideas for guiding them. And checking what AI tells you goes hand in hand with protecting what you tell it: read our guide to AI and privacy.