What are AI hallucinations and can they be stopped?


Weekly Newsletter

Practical AI Strategies

What are AI hallucinations?

Hi everyone,

Since the release of ChatGPT, people have been caught out with large language models' tendency to make stuff up. The industry settled on the word "hallucination" to describe this phenomenon, and I'm sure you've all heard of it, and experienced it yourselves.

If you're a teacher, you've probably had students hand in hallucinated work. If you use AI yourself, you've probably noticed a few wild claims or incorrect references. Hopefully, you haven't been caught out publishing hallucinated work, but if you have, you're not alone!

High profile cases in the news from consulting firms, lawyers, educators and medical practitioners suggest that many people are falling into the trap of AI hallucinations.

This week's article looks at the exact reasons hallucinations occur, and why it's impossible to totally avoid them even if we use careful technical methods. I also explore why the term "hallucination" itself is problematic, and might be adding to people's issues with the technology.

Where does expertise "live"?

If you've been around for a while, you'll have seen quite a few articles on the blog about expertise. It's one of my favourite topics to write about, and I believe that expertise is a vital ingredient in using AI well. Having domain expertise is one of the few things that can mitigate the risk of hallucinations, and it also leads to better outcomes from increasingly powerful models.

But there's a longstanding debate on where expertise "lives". Is expertise something in the head of the expert? Does it leave when the expert exits the room? Or is it something which is conferred upon someone, and a product of society?

That's all for this week!

Cheers,

Leon

PS: If you're a vocational skills teacher, including my Australian VET colleagues, stay tuned for updates coming soon on a new course I'll be putting together at ​https://practicalaistrategies.com/​


Stay informed.
The Practical AI Library curates fresh, practical articles, research, and advice on AI and education—updated every week. Check it out
Check it out → Practical AI Library


211 Tahara Grassdale Road, Grassdale, VIC 3302
Unsubscribe · Preferences

Leon Furze

I'm a educator, writer, and podcaster who loves to talk about artificial intelligence, education, and writing & storytelling. Subscribe and join over 9,000+ educators every week!

Read more from Leon Furze

Weekly Newsletter Practical AI Strategies Practical AI Strategies 2 is off to the typesetters 🎉 Hi everyone, The final of draft of Practical AI Strategies 2 is back from the editor and off to typesetting. That puts us on track for publication in September! Building on the first book published in 2024, Practical AI Strategies 2 covers both the practical and the critical aspects of the technology in five parts: Using GenAI in Education; AI and Assessment; Critical AI Literacy; Beyond Chatbots;...

Weekly Newsletter Practical AI Strategies Hi everyone, First of all, thanks to everyone who responded to last week’s newsletter with their thoughts on “AI;DR” and the trust between readers and writers. I had so many responses that I’m planning a follow up article. It’s definitely a complex topic that a lot of you are interested in! LLMs are Digital Skeleton Keys For years now I’ve been exploring the capabilities of large language models like GPT and Claude. Over time, we’ve seen a shift away...

Weekly Newsletter Practical AI Strategies AI; Didn't Read... Hi everyone, Every Saturday morning, I open up my MacBook after breakfast and write out these emails. If I planned ahead, I could probably write them on Thursday or Friday, but the week always seems to get away on me... Based on the newsletters and emails I'm subscribed to, an increasing number of people are using GenAI to write their weekly emails. I suppose it saves a little time, and maybe makes the planning and scheduling a...