No sense of humour detected: Cardiff research team exposes AI’s biggest flaw

If you were bargaining on ChatGPT writing you a witty Christmas card this year, you’re out of luck


New research by Cardiff Uni has revealed a major flaw in AI’s capabilities: It can’t take a joke.

The research, led by teams at Cardiff University and Ca’ Foscari University of Venice, attempted to counter an argument based on previous research that Large Language Models (LLM) possess similar informational processes to us.

The Guardian reported on the study, which showed that AI had a “fragile” sense of humour.

So, I wanted to see if I’d manage to have any decent banter with a Large Language Model (LLM) if it were a Cardiff University student.

According to Cardiff researchers’ findings, LLMs lack humour, empathy, and cultural nuance. The teams fooled the LLMs by subtly modifying original puns, which exposed AI’s limited capabilities; they could only recall the structure of the pun and failed to actually understand the joke.

In fact, researchers concluded that the LLM’s success rate in spotting puns from ordinary sentences is just 20 per cent. It’s safe to say that AI would not meet Cardiff University’s 40 per cent pass mark.

As AI, particularly ChatGPT, is becoming increasingly popular at Cardiff Uni, I wanted to test if these theories were true, and see if ChatGPT could understand Cardiff-related puns. Now was the time to find out if ChatGPT really has any chat.

Taking inspiration from one of the puns used by the Cardiff University research team, I told Chat GPT: Lectures at Cardiff University really dragon.” Unfortunately, the pun didn’t quite land, as ChatGPT thought we had just mistyped “drag on”. However, after we explained the pun – which is always a sign it’s not that funny – ChatGPT eventually understood. Perhaps this is the start of a lifelong friendship.

Next, in relation to how delightfully clean student areas of Cardiff are, we tested the pun: “It’s important to walk around Caerphilly after bin day because there’s always so much rubbish on the pavement”. This one was slightly more niche so we can let ChatGPT off just this once, but again, the pun fell through. The possibility of a friendship also fell through after we had to explain that “Caerphilly” sounds similar to “carefully”. ChatGPT politely informed us this was “terrible”.

via ChatGPT

To our surprise, our good friend ChatGPT actually understood this next pun without us having to break it down. I told it: “We made some pour decisions in Misfits last night”, to which it caught the wordplay of “pour” immediately. Maybe ChatGPT does have a sense of humour after all, or maybe it’s just a Misfits warrior like most of us.

Some rumours will purport that Cardiff University is superior to Cardiff Met, so we told ChatGPT: “I met someone from Cardiff Meht yesterday”. ChatGPT, once again, failed to get this one and thought, instead, that we were insinuating Cardiff Met had some kind of a drug problem. I clearly did not think this one through… and to set the record straight, Cardiff Met does not have a drug problem, and that obviously wasn’t the intention.

And on that note, my rudimentary research has confirmed one thing: ChatGPT does not have great chat.

For more of the latest news, guides, gossip, and memes, follow The Cardiff Tab on Instagram, TikTok and Facebook.

Featured image via Canva