How to fact-check an AI answer and catch hallucinations

AI chat models write fluent, confident text whether or not the content is right. Sometimes it contains a date that is off by a year, a statistic nobody published, a quote that was never said or a link that leads nowhere. These errors are often called hallucinations, and they are hard to spot precisely because they read exactly like the correct parts.

You cannot switch hallucinations off, but you can make them much easier to catch. This guide explains where they come from, which answers deserve a second look and a short routine for checking them. In Cortex AI the checks happen in one place: a search-capable model, a second opinion and your own source document are all in the same app.

Why AI models make things up

A language model produces the most plausible continuation of your prompt. When it has seen a fact many times, plausible and true usually coincide. When the fact is rare, recent or very specific, the model can still produce something that sounds right, because sounding right is what it was trained to do.

Three situations make this more likely: questions about events after the model's training data ends, questions that need exact figures or citations, and prompts that leave out the context the model would need.

Which answers deserve a second look

• Numbers, dates and statistics. Precise figures without a named source are the most common kind of invented detail.
• Names, quotes and titles. Check that the person exists, said those words and wrote that book or paper.
• Links and citations. Open every link. A plausible URL or reference that does not exist is a classic warning sign.
• Anything recent. Prices, laws, product releases, schedules and news change faster than any model's training data.
• Code you will run. Functions and options that look reasonable may not exist in the version you use.
• Health, legal and money questions. Treat the answer as background reading and confirm it with a qualified professional.

A six-step checking routine in Cortex AI

1. Ask for uncertainty up front. Add one line to your prompt: "Separate what you are sure of from what you are inferring, and say when you do not know." Models generally follow this instruction, and it tells you where to look first.
2. Use a search-capable model for anything time-sensitive. In the model picker, the Web Search filter shows the models that can look things up on the live web through the OpenAI, Gemini or Qwen search integrations. When your plan includes web search, they can search while answering, and the chat shows when a search runs. You can ask for it directly: "Search the web for the current figures."
3. Ask where each claim comes from. Follow up with "Which sources support each of these points? Give me the links." Then open them yourself. A summary of a page is not the page.
4. Give the model the source instead of its memory. If the question is about a specific contract, report or screenshot, attach the PDF or image and ask the model to quote the passage it relied on. Checking a quote against the document takes seconds.
5. Get a second opinion. Send the same question to two models in compare mode, or switch models mid-conversation and ask the new one to review the previous answer for errors. Where the two disagree is exactly where to dig.
6. Confirm with a primary source. For anything you will publish, sign, pay for or ship, check the decisive facts against the original: the official website, the law itself, the documentation, the paper.

Red flags while you read

• Very specific details with no source. "Studies show that 73% of users..." with nothing to point to.
• An answer that changes when you ask again. Tap Regenerate on the same question. If the key fact shifts between versions, the model is guessing.
• References you cannot find. Book, paper or court case titles that do not turn up in a quick search.
• Confident claims about very recent events from a model that did not search the web.
• Agreement with whatever you suggest. If a leading question ("Isn't it true that...?") flips the answer, ask the question again neutrally.

What cross-checking cannot do

Two models agreeing is a good sign, not proof: they may have learned the same wrong claim from the same popular source, and web search can surface low-quality pages as easily as good ones. AI answers can be incomplete, outdated or wrong, and they are not a substitute for professional medical, legal or financial advice. Use AI to find, draft and explain; use primary sources to decide.

Comparison, web search and file attachments depend on your plan, and the app shows which features yours includes. If a response is offensive or harmful rather than just wrong, use Report in the message menu so it can be reviewed.