AI hallucinations: When your chatbot dreams up a three-tailed cat and a donkey-headed surfer

If you thought hallucinations were the exclusive domain of humans, think again. Those clever LLMs, the AI models we’ve getting so fond of, also dream up stuff, and they do so with alarming regularity. Some online sources such as Vectera and GitHub have made it their business to maintain leaderboards of AI hallucination rates, but the range of percentages varies wildly, making it impossible to give one figure for the lot that makes any sense. They also vary depending on the type of task, and that makes it more difficult to pin each AI assistant to an error rate. Sometimes the figures themselves could be hallucinations, but there’s no doubt the average user will bump into a hallucination sooner rather than later.

All you have to do, to see a hallucination in action, is to ask a chat assistant like the Meta.ai that lives in your WhatsApp, to generate an image of something. It won’t take long for you to be startled with a bizarre and hallucinatory take on what you asked. I requested the image of a cat walking over to a door - simple enough - but was promptly presented with a cat proudly sporting no less than three tails. Users have been known to double over in laughter at finding the photo of a friend with double the teeth humanly possible or a surfer emerging out of the sea with a donkey’s head instead of his own, this pastime can get entertaining but creepy quite fast.

AI can give you written and apparently researched facts that are quite confidently wrong. I asked a question that should have led to a straightforward, accurate answer: what time sunrise is today. The chat assistant insisted the sun had already risen, at 7:15am and wished me a good day. It was in fact still 6.05 just then. I went back to sleep. Someone also managed to have a chat assistant agree that 2+2 is 5, so perhaps I shouldn’t complain.

The thing is the AI chat assistants used by the general public for everyday things really hate being wrong. Of course, they don’t have actually emotions, but they still do a good job of expressing them. If you point out a mistake or get uncomfortably specific, you’ll get an immediate apology, some flattery at being so right and fresh information that you won’t know whether to trust or not.

Hallucinations could happen because of inadequate training data. It could be because of a poorly framed request. It could also be from a phenomenon known as ‘overfitting’ in which the AI amplifies or uses irrelevancies and noise. It’s possible that there’s a faulty model architecture to begin with. Guesswork could take place from a lack of real-time access to data, or even from bias or bad actors introducing incorrect data.

The good (or bad) news is there’s no point dumping these AI tools. They’re here to stay and will soon be so tightly integrated with all the technology we use that one won’t be able to weed AI out. That’s already true of smartphone cameras. You have to work hard to shoot a photograph without unsolicited processing done by the AI software in the cameras.

So, there’s no option but to be on the alert for errors. Vary the request. Skirt around the problem with a modified request, even if it means a few repetitions. Try ‘Give me a cat with one tail’. Check with multiple LLM’s, since they’re easily accessible in any case. These tools are very much in development and are improving every day.

There is of course, more cause for worry if the real-world scenario calls for high accuracy. Hallucinations can have devastating results in some use cases. Let’s not even think of anything to do with flying, but unfortunately, hallucinations seem to be more frequent in some professional high-stakes domains. A study by Stanford RegLab showed that in the legal domain, LLMs hallucinate between a range of 69% to 88% of the time for specific legal questions. The use of LLMs has been known to a lawyer quoting non-existent case law when presenting a case. One can only imagine the consequences when the scenario involves medical diagnoses, financial analysis and prediction, or even military scenarios. The very thought is frightening.

So, what do we do with AI chat assistants when they dream up their own reality? Live with them as best as we can. Hallucinations are part of the package for now, but as these systems get smarter, so will their accuracy. And come to think of it, in the grand scheme of things, a cat with three tails might just be the least of our worries.