Short answer
If you have ChatGPT Plus or Enterprise, choose Browse with Bing under GPT-4 and ask it to search the web and link each claim to the page it came from; those links point to real pages it just read. Without browsing, ChatGPT can only produce references from memory, which may be inaccurate or entirely invented, so treat them as leads to check, never as proof.
Follow-up question Why must you still verify them?
Asking ChatGPT for sources is a standard prompting tip, and A Beginner's Guide to ChatGPT includes it: when you need evidence, ask for references that support each claim. What has changed this fall is what you get back. Since September 27, browsing is back for ChatGPT Plus users, and on October 17 OpenAI moved it out of beta. With browsing on, ChatGPT can search the web and give you direct links to the pages behind its answer. With browsing off, it is still writing from what it learned in training, and its "sources" are a very different thing.
Understanding that difference is the whole skill. A citation is a pointer to where a claim came from: an author, a title, a date, ideally a link. Citations from a browsing session point to pages the tool actually opened. Citations from memory are text the model generated because citations usually look like that.
How to turn browsing on
In ChatGPT Plus or Enterprise, open the model menu, pick GPT-4, then choose Browse with Bing. You no longer need to switch on a beta setting first. Free accounts using the standard model don't have browsing, so the steps below for "browsing off" apply to them.
Browsing also lifts the old limit on recent events. OpenAI says the browsing model is no longer restricted to data from before September 2021.
The same request, two ways
Here is a prompt built to get checkable citations:
Search the web and answer: what are Canada Revenue Agency's rules for claiming
a home office expense as an employee this year?
For every factual claim, add the link to the page it came from.
Prefer official government pages. If you can't find a source for a claim, say so.
At the end, list the pages you used with their titles.With Browse with Bing on, a good reply shows that it searched, then answers in short paragraphs with links after each claim, ideally to canada.ca pages. You can click each one.
With browsing off, the same prompt still produces an answer, and it may still produce something that looks like a reference list. But the model can't open pages. The titles and links are reconstructed from memory: some will be right, some will point to pages that have moved, and some may not exist at all. Change the request when browsing is off:
Answer from what you know, and tell me which points I should verify.
Instead of links, name the type of official source I should check for each point
(for example, the CRA page on employment expenses) so I can look it up myself.This version doesn't invite the model to invent URLs. It turns ChatGPT into a guide that tells you where to look, which is a job it does well.
A two-minute citation check
Whatever the mode, run every source through four questions before you rely on it:
| Check | What to do | Red flag |
|---|---|---|
| Does it exist? | Click the link, or search the exact title | Error page, or no result for the title |
| Is it the right page? | Find the claim on the page itself | The page covers the topic but never says that |
| Is it current? | Look for a date or "last updated" note | Last year's rules presented as this year's |
| Is it trustworthy? | Check who publishes it | An unknown blog cited for a legal or medical point |
The second check catches the most problems. A real link next to a sentence does not prove the sentence came from that page. Use your browser's find command (Ctrl+F or Cmd+F) to search the page for a key number or phrase from the answer.
Try it: Ask ChatGPT a question you already know the answer to, such as a rule from your workplace or a fact about your town, and request a source for each claim. Run every source through the four checks above and count how many pass all four.
Why must you still verify them?
Because a citation is only as good as the page behind it, and ChatGPT cannot tell a true claim from a well-worded one. This spring gave a hard example. Lawyers suing the airline Avianca filed a court brief citing six earlier cases that did not exist. One lawyer said he had found them with ChatGPT and had checked them by asking the chatbot whether they were real. It said yes, and even claimed they could be found in the legal databases Westlaw and LexisNexis. Judge Kevin Castel wrote that the cases appeared to be bogus decisions with bogus quotes. Asking the same tool to confirm its own work is not a check.
Browsing lowers the risk but does not remove it. ChatGPT can still summarize a page loosely, mix up two sources, or attach a link to a claim that came from somewhere else. The web pages themselves can be outdated or wrong. And the model chooses which few pages to read, which may not be the best ones.
So treat ChatGPT's citations like a research assistant's notes. They save time by pointing you in the right direction, and when browsing is on they usually point somewhere real. But the decision to trust a fact is yours, and it should rest on a page you opened and read yourself.
Sources
- ChatGPT — Release Notes — OpenAI Help Center, October 17, 2023
- A lawyer used ChatGPT and now has to answer for its 'bogus' citations — The Verge, May 27, 2023
- A Beginner's Guide to ChatGPT — Holistic Hybrid, April 18, 2023