The Seillans Festival of Lights; courtesy of DALL·E 3 AI
I am not going to be writing often. Too busy "doing AI" to write about it. Lol. But I will try to hold to quality over quantity, and my usual irreverent take on the field.
What is Hallucination?
In my day job, we work quite a bit with IBM (in tackling hallucination, biases and other AI governance challenges). So I am going to reference their own preferred definition:
AI hallucination is a phenomenon wherein a large language model (LLM)—often a generative AI chatbot or computer vision tool—perceives patterns or objects that are nonexistent or imperceptible to human observers, creating outputs that are nonsensical or altogether inaccurate.
IBM go on to talk at length about the implications of the problem and to consider what we need to do to "fix it". As they put it, users expect "a correct answer to a question", not some made-up garbage. An example they reference is Google’s Bard chatbot incorrectly claiming that the James Webb Space Telescope had captured the world’s first images of a planet outside our solar system
Indeed, all the body-of-knowledge on this is in 100% agreement. Hallucination is a big problem and presents a major barrier to gaining people's trust in AI and driving adoption. Worse, it is hardly rare. Vectara, a start-up founded by former Google employees, found in their research that even the best chatbots hallucinated at least 3% of the time - and the worse up to 27% of the time!
What causes hallucination?
A cynic would describe an LLM as just a fancy predictive-text engine. Yes, like that annoying feature on your i-Phone that most of us turn off. The core functionality is indeed to predict (using probability) the next-most-likely output in a string of text or code, based on the data on which the LLM has been trained. As Brent Mittelstadt, of Oxford University, puts it, “there is no absolute requirement for them to be factual, accurate, or line up with some sort of ground truth”; they do not inherently know fact from fiction.
But what is the worst that could happen?
You may well ask! Google's new artificial intelligence (AI) search feature - "AI Overviews" - has attracted some unwanted media attention recently for advising users to make cheese stick to pizza better by using "non-toxic glue", and for saying "geologists recommend humans eat one rock per day".
Investigations suggest these answers may derive from comments on Reddit or satirical articles on The Onion. In other words, if you train the AI on data of humans being funny or ironic, what do you expect?
Whilst these instances are certainly isolated and mostly just amusing, Google's AI Overviews are causing concern - particularly in the medical community. Research from SearchEngineLand suggests AI Overviews appear far more frequently - up to 63% of the time - in Healthcare-related search queries. So imagine what would happen if people were searching for cures for cancer and the like, whilst using an LLM trained on Reddit. Or Gwyneth Paltrow's Goop.
My Epiphany
Folks, I had an epiphany the other day. I think it's probably that longstanding right-brain disorder of mine again - or an undigested piece of cheese - but all the same, I wanted to share it. I was using ChatGPT, as you do, to write a guest blog article to promote a Travel Blog of mine on events in Provence and things to do on the French Riviera. Yes polluting the internet, I know, but if you can't beat 'em, join 'em, right?
I made a classic prompting error of the sort that frequently produces AI hallucination. I asked it to generate a post covering five summer festivals in the Pays de Fayence; a canton in the Var with a population of just 29,000 people. Asked to work on such a narrow data proposition, ChatGPT returned four events just fine but struggled for a fifth. So it concocted "The June Festival of Lights" in Seillans; complete with pagan bonfires, torches, and all manner of embelishments. It sounded fantastic but - as a local - I sadly immediately recognised it as "fake news".
But here's the thing. The more I thought about it, the more I thought "what a great idea for a festival". I could absolutely see why ChatGPT had generated it. Seillans is one of the best-known and most beautiful villages in France. Scary Spice used to live there, I think [citation needed]. But it doesn't have a proper big-name summer event. It should have! And there is definitely a bit of a hole in the local social calendar around that time, which feels like it shouldn't exist. Plus, pagan stuff around the Solstice? Of course! Fires and torches? Ace. I even started writing an email to the Mayor, to suggest he consider it as an idea.
If pigs should fly
Anyone who uses AI frequently will have experienced a moment like this. From my time testing BingAI, I remember one chap generating an itinerary for a trip to Mexico City; where it invented an entire restaurant (complete with menu and website link) that did not exist. I also recall a person seeking a reading list of new books (based on their interests) where it generated a Book Title, Author Name, and Synopsis for - yes you've guessed it - a book that hadn't been written. It's not so very hard to reproduce this yourself if you are curious. Just like Dileep Mishra did recently in his "AI Hallucinations or Bullsh*t?" article on LinkedIn.
The funniest aspect? Both individuals registered significant disappointment at the realisation they could not savour that nosh or nose through that book. Both would have 'filled a gap'. Why? Because, probabilistically speaking, both should have existed. They were logical. Just like my Festival in Seillans. Which was why ChatGPT generated them in the first place.
The epiphany came in two parts. Part one was a realisation that ChatGPT could be about the best new business idea generator ever invented. I am pretty sure that if you went and built that restaurant in Mexico, that people would come visit it in droves. The menu sounded yumm and it was in a part of the city with good footfall but lacking that kind of establishment. Get my drift? And that festival in Seillans would be a success for sure.
Part two was a slower burn. I came to realise that hallucination is really just a rudimentary form of imagination. Scientists generally agree that the emergence of imagination in humans served an important evolutionary purpose; allowing us to conjure up multiple hypothetical futures (emerging from the present), to help us decide which choices we should make; making survival for longer more likely in complex or hostile environments.
So, what if we are getting this all wrong?
So, if hallucination is really a basic form of imagination, and it's results akin to creativity or invention, then why on earth would we want to eliminate it? Why are we so determined to lobotomise AI before it's even crawled onto land?
If you know me a little by now, you will know that I just can't resist the temptation to personify AI and to see the flickering of real intelligence where others just see predictive text. So, would it surprise you to know that scientists still don't really understand (yet) how our brains distinguish between what is real and what is imagined; other than observing we struggle more to do so in the period just after we wake up from a long sleep? Recent research suggests it may be more about relative "strength of signal" than some separate executive function in - or region of - the brain.
Perhaps once we have figured out how we humans can distinguish hallucinations, we might be better able to do so with AI. Either way, I would rather we work with imagination than work to eliminate it!
Final thoughts
Recently, I was delighted to note that Perplexity.ai thought my holiday home in Provence was the "best luxury villa in Montauroux". It correctly referenced the private pool, panoramic view, modern air-conditioning, and fast WiFi. Of course, it hallucinated the hot tub altogether. But perhaps it really should be there. Yes, I think it will be.
© David Viney 2024. Licensed for re-use with attribution under CC BY 4.0
This article is the original & canonical blog post. For re-prints & syndication, my preferred attribution code snippet is:
<a href="https://ecency.com/hive-196902/@dviney/what-if-ai-hallucination-was">
What if AI hallucination was imagination?</a>, by
<a href="https://david-viney.me/">
David Viney</a>, under
<a href="https://creativecommons.org/licenses/by/4.0/">
CC BY 4.0</a> attribution license.