Researchers have developed an AI that is so good at writing text they've decided to keep the technology behind it secret over fears it'll be exploited to write high-quality fake news.
"Due to our concerns about malicious applications of the technology, we are not releasing the trained model," the research company OpenAI wrote in a blog post on Thursday.
The AI, called GPT-2, can produce text of "unprecedented quality" when compared with other computing systems by simply giving it a short writing prompt. For instance, the AI can pump out fiction. It can also finish a homework assignment. All you have to do is serve up the topic or the beginning first sentence, and the AI will do the rest.
Although the results don't always make sense, the AI's writing can be very human-like and natural sounding about 50 percent of the time. "The orcs' response was a deafening onslaught of claws, claws, and claws; even Elrond was forced to retreat," it wrote in attempt at Lord of The Rings fan-fiction. (You can find more examples here.)
So how does it actually work? "GPT-2 is trained with a simple objective: predict the next word, given all of the previous words within some text," OpenAI explained in its blog post. To predict the next word, the AI was trained to model its writing from a dataset of 8 million webpages that Reddit users had posted links to.
As a result, the AI can also change its writing style based on the prompt you give it. For example, if you feed it the opening line to George Orwell's sci-fi novel Nineteen Eighty-Four, the system will continue to write the remaining text like it were authoring a dystopian work of fiction.
The AI is also adept at writing news, whether it be real or fake. Give it a writing prompt in the style of a news article, and the AI can convincingly flesh out the rest, even if you tell it to write about scientists discovering unicorns.
"The scientist named the population, after their distinctive horn, Ovid's Unicorn. These four-horned, silver-white unicorns were previously unknown to science," it wrote. The same AI can also be used to write a human-like rant on why Recycling is bad.
The examples underscore why OpenAI — an Elon Musk-backed research company that's focused on developing a safe path to artificial intelligence — has decided to keep the GPT-2's computing coding largely secret. They're concerned the AI could fall into the wrong hands and pump out online hoaxes, fake social media posts, and email spam on a wider scale than any human could do.
Of course, the same technology also has potential benefits. Imagine AI writing assistants, more natural-sounding chat bots and better language translation. But for now, OpenAI has decided to only make available a smaller computing model based on GPT-2 that the public can experiment with. However, none of training code for the AI or the data it was modeled on will be released.
"We need better policies around the distribution of increasingly transformative tech," said OpenAI's policy director Jack Clark in a tweet. "We're not really prepared for a world where this stuff gets incredibly cheap."
The norms over how AI technologies should be made public will need to be hammered out over time by the research community, OpenAI added in its blog post. To prevent potential abuse, the research company is also advocating that governments step in and consider expanding ways to monitor the impact and spread of AI technologies.