![](https://static.standard.co.uk/s3fs-public/thumbnails/image/2018/05/29/10/jimmy-wales04.jpg?width=1200&auto=webp)
By now, even those of us who live under a rock have become familiar with ChatGPT, the AI-powered chatbot that can find us the answer to almost every question under the sun. It seems to be able to effortlessly write reports, compose letters or even poetry — for any subject its asked about, it dutifully complies and does so at breakneck speed. It has even been known to declare its love for the users that interact with it.
The technology, developed by OpenAI — which execs at Microsoft have found so impressive, they’ve bought a $10 billion stake in — scours the net to piece together a vast corpus of human-made literature, and draws on this to find users the answers to users’ questions it thinks are most plausible. In this respect, as one journalist put it, the software is little more than a ‘spicy autocorrect.’
But now imagine the roles are reversed. Instead of AI bots drawing on a body of human writing, imagine that in the future, most of the material we humans will read about will be written by AI. Such a move could send shockwaves through the world of publishing, newsgathering and social media, and it could be just around the corner. It’s a future that Wikipedia founder Jimmy Wales is already spending a lot of time thinking about, as he weighs up how the world’s biggest online encyclopaedia will evolve in the years to come.
“The discussion in the Wikipedia community that I’ve seen so far is…people are cautious in the sense that we’re aware that the existing models are not good enough but also intrigued because there seems like there’s a lot of possibility here,” Wales said.
“I think we’re still a way away from: ‘ChatGPT, please write a Wikipedia entry about the empire state building’, but I don’t know how far away we are from that, certainly closer than I would have thought two years ago,” he said.
One of the issues with the existing ChatGPT is what they call in the field ‘hallucinating’ — I call it lying.
Wales says that as much as ChatGPT has gripped the world’s imagination over the past few weeks, his own tests of the technology show there are still plenty of flaws.
“One of the issues with the existing ChatGPT is what they call in the field ‘hallucinating’ — I call it lying,” he said.
“It has a tendency to just make stuff up out of thin air which is just really bad for Wikipedia — that’s just not OK. We’ve got to be really careful about that.”
Wales is not the only person to have spotted ChatGPT’s ability to hallucinate.
Last month, David Smerdon, an economist at the University of Queensland, asked ChatGPT: ‘What is the most cited economics paper of all time?’
Within seconds the AI chatbot came up with an answer. “The most cited paper of all time is “A Theory of Economic History” by Douglass North and Robert Thomas, which was published in The Journal of Economic History in 1969.”
The answer sounded perfectly plausible. But when Smerdon went to find the paper it had referenced, to check it was as popular as ChatGPT claimed it was, he found that no such article existed. The authors were real, published writers, and the journal was real too. But the name of the paper was completely fabricated. Citation needed.
But beyond failing to tell the truth, Wales says the technology often fails to spot the internal contradictions in its own writing.
“I asked ChatGPT, did an airplane crash into the empire state building? And it said no, an airplane did not crash into the empire state building,” he said.
“But then it went on to say, there is a famous building in New York and one of the most famous things that happened is when a B25 bomber crashed into the empire state building.
“I said, is a B25 bomber a type of airplane? It said says —so I said you were wrong when you said a plane didn’t crash into it.
“And it said, you’re right, I apologise for my error.”
But while full AI authorship is off the cards in the near-term, there’s already plenty of discussion at Wikipedia on what role AI technology could have in improving the encyclopaedia in the months ahead.
“I do think there are some interesting opportunities for human assistance where if you had an AI that were trained on the right corpus of things — to say, for example here are two Wikipedia entries, check them and see if there are any statements that contradict each other and identify tensions where one article sems to be saying something slightly different to the other,” Wales said.
“A human could detect this but you’d have to read both articles side by side and think it through — if you automate feeding it in so you get out hundreds of examples I think our community could find that quite useful.”
One of the criticisms levelled at Wikipedia over the years is the scope for bias in both the way its volunteers describe particular subjects as well as the way they choose which topics deserve more or less prominence. The overwhelming majority of Wikipedia’s millions-strong army of volunteers are male and white, meaning they inevitably — even if inadvertently — interpret world events with a male-centric and white-centric perspective – and may not bother covering important topics at all if they don’t see any merit in doing so. Could AI be a way to counteract some of these biases, and make Wikipedia a truly neutral, impartial source of information once and for all? Wales is not convinced.
“We know a lot of AI work has run into bias very quickly, because if you train an AI on biased data then it’s just going to just follow that bias, and a lot of people in the AI world are focused on that problem and they are aware of it,” he said.
In other words, there’s a real chance that an AI-powered Wikipedia won’t put a stop to biases, but exacerbate and perpetuate them. But where it could be most useful is in spotting things that are missing the encyclopaedia’s coverage – analysing all the available information in the world, pairing bits of it up with corresponding Wiki entries and identifying the gaps. And there might be a lot of them – leading to a ballooning of Wikipedia content. But Wales is unfazed by this.
“Using AI to triple the number of Wikipedia entries wouldn’t increase our running costs by more than £1,000 a year,” said Wales.