Smart text-to-speech technology can take advantage of a strange viral phenomenon in online communities: people voting on stories online, nearly all of them bots (a term we’ve all heard a few times).
Researchers now say that their AI created fake content that grew tens of thousands of times faster and longer than bots that get reported as fake material.
In a paper published today in the journal Proceedings of the National Academy of Sciences, researchers show that the AI can generate news stories, delivered via text, at speeds ten times faster than bots — because, well, it uses artificial intelligence.
Overall, the AI produced news stories which typically spanned more than 1,200 words and used unique formatting. And it created headlines of even longer length, averaging roughly 100 words.
The latter is something that would be pretty difficult for bots to come up with. Traditional spam pieces of content — i.e., anonymous posts and links — have their own style, and sometimes they contain outdated information. In addition, bots tend to resort to embellishments.
“In the early stages of newsworthiness, determining which social elements and specific content is reliable is a complicated and subjective endeavor,” study co-author Yimin Li, a postdoctoral researcher at the Social Media Research Laboratory of the University of Southern California in Los Angeles, told VentureBeat. “Our work shows that personalized keyword search can rapidly establish such determinations.”
This paper shows that the AI developed the news content without any bots in its system. But the AI still had to study dozens of stories to find the ones most relevant for interpretation.
Li’s team, including the Social Media Research Laboratory’s Murray Douglas, used machine learning techniques to identify the keywords included in the tweets. They then examined the content of the stories posted on the social media site Twitter before and after the algorithm selected articles from a database of over 6 billion tweets. After analyzing the published and retweets of the articles, the AI correctly identified some 5,000 news stories on Twitter and 4,500 on Facebook.
In the end, the AI generated headlines which ran on top of stories with similar articles. Instead of using bots, the AI used a highly accurate algorithm that analyzed the tweets of a high number of people, so that the AI could more accurately identify significant stories.
“Our search AI was human-generated,” Li said. “Our algorithm was not human-driven. Even though the AI was human-generated, our algorithm was closely supervised by humans to ensure accuracy.”
The paper specifically discusses guidelines for applying bots to stories and determining where to build content and how to ensure the accuracy of the article. While the AI didn’t specify how to categorize stories into “news” or “tweets,” it did inform users of the articles they should interpret as news. For example, the bots will turn a question into a trending trend if they know, from the data, that it’s a top topic on Twitter.
All of this may be fine if you’re an advertiser, but Li’s team argues that automated content creates false news and leads to engagement that is distracting and unreliable.
“We found that bots are statistically less reliable in making socio-economic determinations (e.g., income), suggesting that bots are less reliable at manipulating social content,” the researchers write. “We can better understand how bots can mislead users and help them better understand news content.”
“We also saw that by looking at social content, the AI algorithms are able to establish consensus rankings of news content more easily,” Li added. “They also use more accurate statistical methods to assess the validity of news stories.”
Interestingly, the AI focused primarily on 60-second articles in newspapers. The AI suggests that publishers should focus on shorter news articles.