Why does ChatGPT sometimes make things up?
Understand why a fluent reply may still need a fact check.
ChatGPT builds replies from patterns and the context and tools it can use. That can produce a plausible answer without enough support. People often call invented AI facts 'hallucinations'.
It can use web search in supported settings. Searching does not remove all errors: it can pick a weak source, misread it or claim a link says more than it does.
Ask it to mark unknowns, then check important claims yourself. A request not to guess is useful, but it is not a guarantee.
- Try the sample below, where one fact is known and another is missing.
- The reply should keep Tuesday as the day and say the room is unknown.
- If it invents a room, point to the source and ask for a correction.
- For a real task, open the original source. Check the date, context and exact claim before you use or share it.
Practice only. A club notice says: 'We meet on Tuesday.' Which day and room does it give? Use only the notice. Mark any missing fact as unknown.
Paste the message into ChatGPT, then fill in any blanks before you send it. ChatGPT opens in a new tab. Opening it does not send the message.
A real source may exist and still not support the claim. Check what it actually says.
Sources and more help
These links open in a new tab.