Does ChatGPT make up references?

A 2026 Royal College of Surgeons of England study checked 1,249 references produced by AI chatbots. For some of the chatbots, between a quarter and a third of the references were made up or couldn't be found. A BMJ Open audit the same month asked five chatbots to back up health answers with references. Not one gave a fully accurate list. The median reference list was only 40% complete.

This isn't a rare glitch. It's how these tools work. They produce text that looks like a reference list, because they've seen millions of reference lists. Looking right and being right aren't the same thing, and the tool can't tell the difference.

Why the fakes are hard to spot

At a glance, a made up reference looks completely normal. Author names that sound real. Journal titles that are real. Sensible years. Something that looks like a DOI. Nothing about it tells you the paper doesn't exist.

That's what makes this risky in clinical work. A wrong dose looks wrong when you check the BNF. A fake reference looks fine until you go looking for the paper, and most people don't go looking.

I found this out myself. For Issue 04, I asked ChatGPT for community services for older adults in Newham. It gave me a confident, tidy list. One of the links went nowhere. Nothing about the answer warned me. The only reason I knew was because I clicked it.

How many of us click every link?

They also give themselves away in the same way every time. The details are specific, but you can't check them. A real reference gives you something to hold: a DOI that works, a PubMed entry, a PDF. A fake one gives you the look of those things with nothing behind it. Once you've caught one, you start noticing them everywhere.

The two minute citation check

Before you act on, share or save anything an AI tool has referenced, do this.

1. Search the exact title. Put the title into Google Scholar or PubMed, in quotation marks. A real paper shows up straight away. A fake one returns nothing, or a real paper with a different title, which means the AI has stuck real authors onto a title it made up.

2. Check the DOI works. Paste it into doi.org. If it takes you to a different paper, or nowhere, don't trust the reference.

3. Open it and read it. Even when the paper is real, check it actually says what the AI claims. After outright fakes, the most common problem is a real paper being quoted for something it never said.

Two minutes. Three steps. That's it.

Try it this week

Next time an AI tool gives you references, pick one and check it before you read anything else. Time yourself. Notice how differently you read the rest of the answer once you know whether it holds up.

If you want to practise, the "Would you send it?" exercise on this site hides a fake journal reference inside a clinic letter. See if you spot it before the reveal.

Before you rely on this

These studies looked at AI chatbots in general. Newer tools, and ones that search the web before answering, may do better, and the numbers will keep changing. The habit doesn't depend on the numbers. However good the next tool is, checking still takes two minutes.

This is general information, not academic or clinical advice. If you're publishing or writing something formal, follow your institution's rules on citations.