Computers might be close to being able to write better than humans, but not without dropping some real clangers.
Recently developed AI technology like Generative Pre-trained Transformer 3 (GPT-3) technology is being hailed for writing original prose with human fluency.
The Open University Emeritus Professor Michael Sharples has been researching the application of AI to writing for 45 years and co-authored the upcoming book Story Machines: How Computers Have Become Creative Writers.
He tells Kim Hill that AI has reached a dangerous point where computers can generate convincing text, but the software lacks common sense and moral purpose.
Photo: 123RF
“The problem with these AI writers is that they are incredibly convincing,” Sharples says. “They write really plausible text, they’re general purpose language machines, they can answer questions, they can translate, they can summarise scientific texts, but they haven’t had any experience of the world.
“They don’t know how to make judgements about what’s legal, honest, decent.”
GPT3 technology can have many uses, but its fundamental weakness is that it doesn’t have an internal system that knows how to reason, he says.
“My colleague who co-authored the book Story Machines with me has written a very different sort of AI programme that can reason about its output but it’s much harder to craft that and it’s not so versatile, it’s not such a great wordsmith.
Photo: supplied
“Some time in the future, we might be able to put machines that reason about their output together with machines that are superb at creating convincing text, but we haven’t done that yet.”
(You can try Sharples and co-author Rafael Pérez y Pérez’s GPT-3 story generator online here.)
Although GPT3 isn’t used that frequently now, Sharples believes it will gain traction as big companies pick up on it and with the Open AI company making the interface publicly accessible to all.
“There are start-up companies now that are setting up that will use that technology to help writers in various ways, some quite nefarious ways like that company to help students write their essays, their academic assignments, but also to help authors, to help newspaper writers, journalists. So it’s not big at the moment but it will be over the next few months and years.
“Google has developed one … and Meta, they’ve incorporated it now into facebook, even the Chinese, they use this system call Wu Dao 2.0, which claims to be the most powerful story generator, text generator in the world.
“They [GPT3 technologies] are going to be controlled by big companies and how they will be used is not just for writers; Microsoft invested in it because it would generate computer programmes automatically. So, there are many different applications.”
Some developers have added filters that will flag any rude words but there’s still a risk these technologies could absorb implicit bias because it trains on billions of pieces of data online, Sharples says.
“You would think that the more powerful these machines got, the more data they were trained on, the better they would be, the more responsible they would be, but actually it’s the opposite.
“There was a paper that was literally published yesterday that has looked at these models and tried to see what happens when you can scale them. And when they get bigger, then they become worse in their social bias, they’re more racist, they’re more sexist and that’s because they pick up the subtleties of the internet.
“It’s going to be very difficult, because you’d have to train them in a different way, you’d have to train them on non-sexist language, you’d have to train them on non-racist internet pages and it’s very hard to do that.”
While Sharples believes GPT3 is a genuinely creative tool that can be useful for writers when used wisely, he is also concerned about its impact on education.
“I’m excited because it’s an enormously powerful technology and it’s one that’s creative. Previous technologies were kind of assistive, they might help you with your spelling, [but] this is genuinely creative.
“As an educator, I’m worried that students are going to start just using them to write their essays.
“Already students are doing this, and universities don’t realise, schools don’t realise, so it’s going to change the way we think about education. So that’s the worrying side of it, or possibly if we have to rethink education, if we have to rethink assessment, maybe that’s a good thing.”
GPT4 is already on the horizon, and it will be capable of putting text with pictures and video, Sharples says.
“So you will have computer-generated soap operas, computer-generated fantasy movies for you, so that’s I think the next stage and that’s already coming, starting to put together images and text, and then movies will come along later.
“You will get immersive computer games where the characters are generated by computer and you can immerse yourself, you can involve yourself in these fantasy worlds, that’s why Meta, the company that owns facebook, is so interested in these, because it wants to power the metaverse with these sorts of AI programmes.”
