Fake links, made-up sources, clauses that aren't there: how to catch AI when it invents things

On the OpenAI community forum, users have filed detailed reports of ChatGPT inventing web links and article titles, even in web-search mode and even after being asked to verify them. Another thread from May 2025 noticed one model adding hallucinated sources to the end of answers "very frequently".
It isn't only casual users. In a Chinese developer forum, someone testing legal AI tools in June 2026 found one claiming a contract had a force-majeure clause in section 5. The contract had no such clause.
Why AI makes things up
A chatbot isn't looking facts up in a database. It predicts what text is most likely to come next. When the honest answer is "I don't know", a plausible-sounding guess often comes out instead, in the same confident tone as everything else. Names, numbers, citations and URLs are where this happens most.
Three ways to get fewer made-up answers
- Give it the source. Don't ask it to remember an article or contract; paste it in and ask it to quote the exact sentence each point comes from.
- Give it permission not to know. Add: "If you're not sure, say 'not sure'. Don't guess. List your sources."
- Use a tool that shows sources, like Perplexity, for factual questions, and click through to at least one.
Always check these yourself
- Numbers, dates and amounts
- Names of people, books, companies and laws
- Every link, especially if you're going to share it
- Anything legal, medical or financial
The more confident the answer sounds, the more worth checking it is.
Step-by-step page with copy-ready prompts: AI made something up: how to catch it
Sources: OpenAI community threads Hallucinated URLs + Fake Article Titles in Web Mode (2025-05-07) and 4.1 prone to hallucinating sources (2025-05-26); V2EX: legal AI inventing contract clauses (2026-06-16, Chinese). Checked 2026-11-12.
评论
发表评论