G’day, it’s AB.
Hallucinations. It doesn’t sound like a good word – and at the extreme, it’s not.
But I’m here to tell you, it’s not all bad.
Large language models need to think creatively. You don’t always want a model that just gives you the facts, or offers a direct, literal translation from one term to another. You want a bit of flair.
In the field, that flair is often called heat – and you can turn up the temperature. Turn it up too much, though, and things can get chaotic. That’s when things start to go a little strange.
So – how do you spot a hallucination?
It’s usually when something just doesn’t look right. To (badly) misquote a Marvel character:
“Sometimes you end your sentence with the wrong banana.”
Sometimes it’s obvious. And often, just asking the same prompt again will give you a completely different result.
The model might tone it down on the second try. Like I mentioned in a previous video, rephrasing your prompt or asking for a bit less can be enough to smooth things out.
Another way to spot hallucinations is by looking for missing links or references. If you’ve run a web search, or if you’ve uploaded your own files into the chat, you should see references pop up – underlined links you can click. These will take you to the source: a paragraph or sentence from the original document or site.
If you don’t see those, and the response still sounds a bit weird or overconfident, you can dial it back. Just prompt it to be less creative, less risky.
But ultimately, the responsibility lies with you.
At the bottom of almost every LLM or chat interface, you’ll see the same disclaimer:
“This model may produce inaccurate information.”
So – your job is to stay alert.
Here are three things you can do:
First, spot the obvious weirdness. Then retry. Maybe soften your own prompt a bit.
Second, follow any links. Make sure they’re real – and that the content the model describes actually appears on those external sites or documents.
Third, tell the model how you want it to behave.
You can say: “For this chat, I want you to be professional, clean, concise.”
Or: “For this chat, I want you to be fictional, creative, out there, zany.”
It’s up to you.
A hallucination isn’t ideal – but ironically, you do want a little creativity, a little flair. Without it, these tools wouldn’t be nearly as much fun.
Hope that helps!