AI Chatbots Fact-Check Each Other in Fight Against ‘Hallucinations’

San Francisco, – As artificial intelligence (AI) chatbots get better and better at sounding like us humans, there’s a kinda funny but scary problem: they sometimes spit out totally bonkers info. We’re talking straight-up “hallucinations,” like a robot seeing unicorns. But now, researchers think they’ve found a fix – and get this, it involves using other chatbots to play AI detective.

Can Chatbots Really Fact-Check Each Other?

A super interesting study just dropped in the fancy science journal Nature, and it’s got everyone talking. It says we can use popular chatbots like ChatGPT (you know, the one everyone and their grandma is using) and Google’s Gemini to basically become truth-tellers, sniffing out the BS from their AI buddies.

This wild idea comes from Sebastian Farquhar, a computer wiz over at the University of Oxford. He and his crew are all about using the power of “large language models,” or LLMs, which is just a fancy way of saying the brains behind these chatbots.

The Science Behind the AI Showdown

So, picture this: LLMs are like those friends who’ve read every book ever. They gobble up massive amounts of text and code, learning to write like a human by guessing the next word, kinda like a super-powered autocomplete. Problem is, they’re all about patterns, not actual understanding. That’s why they sometimes say stuff that’s, well, totally out there.

Farquhar’s big idea is to make the chatbots debate each other, like a digital showdown. One chatbot spills the tea, answering questions or writing stuff, while another one acts like a skeptical detective, looking for any fishy business. By asking the same question a bunch of times and comparing the answers, the detective bot can be like, “Hold up, something ain’t right here…”