In the battle for AI supremacy, the best defense is a good offense.
That’s the philosophy behind Grover, a neural network application that is fighting disinformation by creating its own fake news.
“We have to be one step ahead of potential misuse,” said Yejin Choi, a University of Washington professor and lead author on the project, which was a collaboration by researchers at UW and the Allen Institute for Artificial Intelligence (AI2).
Choi said that while today’s fake news is mainly written by humans, the age of fake news created by neural networks is just around the corner. In that environment, the people with the best fake-news generator also have the best fake-news detector. Grover’s approach is similar to that used by cybersecurity professionals, who regularly attack their own systems in order to discover weaknesses.
Grover’s remarkably realistic news-making engine mimics the style and tone of specific publications and authors. The detection tool was able to tell the difference between human-written news and machine-written news 92 percent of the time, researchers said. Rather than looking at a single word that may be out of place, the model identifies statistical patterns in the overall text to decide if a piece of writing is fake or not.
The researchers made the generator open to the public. We put it to the test by asking it to create a GeekWire article with the headline “Microsoft reportedly in talks to purchase Nintendo,” and the resulting article was eerily realistic.
Here’s what Grover came up with:
Microsoft reportedly in talks to purchase Nintendo
May 29, 2019 – Nat Levy(Fake article)
Bloomberg reported Wednesday that Microsoft is in talks to buy Nintendo for about $60 billion in cash.
Microsoft’s shares rose 4.6 percent to $122.96, closing at the highest price since December 31, 2017. Nintendo’s stock was also up 5.9 percent to $119.84. Both jumped more than 10 percent.
A deal between the companies is reportedly being discussed for between $120 and $130 per share. Sources tell Bloomberg that talks could fall apart at any time.
Representatives for Microsoft and Nintendo could not be reached immediately for comment.
[Update: CNBC reported that the two companies are in a private, high-level negotiation, with the $30 billion ask on the table and the other asking for at least $100 billion. Bloomberg said that reported is not accurate.]
Nintendo is famed for its role in the video game industry with its arcade and game systems, including the company’s Switch console, which has racked up sales of more than 50 million units worldwide. A deal could raise questions about whether Microsoft’s vision for gaming centers on software and other digital means, instead of physical consoles like Nintendo’s hardware.
The two companies have not collaborated on a video game in decades, as Microsoft typically prefers to stay away from hardware and instead relies on software, its Xbox gaming console being a classic example. The partnership would, however, put Nintendo’s iconic characters in Microsoft’s Xbox One and Windows products, as the Xbox recently grabbed the crown of the top console in the U.S. for the first time.
Nintendo created and publishes some of the most successful video games, and its characters are as loved as they are recognizable. Mario, Zelda, Donkey Kong and Link are among Nintendo’s most popular character franchises. The company has one of the top ten best-selling home consoles of all time.
It will be a tough sell for Microsoft, however, as the company is also focused on developing games and building a gamer-first approach to its technology products, with games particularly at the center of what it does. For Nintendo, it will keep the focus on game development instead of hardware, as its Wii U and 3DS consoles continue to flounder.
Tools like Grover could become essential to social media platforms as fake news more automated and sophisticated — a serious concern as the 2020 election season ramps up. Just this week, Twitter purchased London-based startup Fabula AI, which spots fake news based on how those articles spread online.
In their published paper, the researchers write that the threats posed by adversaries seeking to spread disinformation “are real and dangerous.” They recommend that news media companies employ similar neural network tactics used by platforms such as YouTube to detect fake news.
“An ensemble of deep generative models, such as Grover, can analyze the content of text — together with more shallow models that predict humanwritten disinformation,” they wrote. “However, humans must still be in the loop due to dangers of flagging real news as machine-generated, and possible unwanted social biases of these models.”
Researchers at OpenAI, an AI research nonprofit backed by Elon Musk, earlier this year refused to release its own text generator, GPT-2, because of “concerns about malicious applications.” Some critics complained that the decision violated the organization’s vow to make its work public. OpenAI later released a watered-down version of its tool at TalkToTransformer.com.
Going forward, Choi said that more collaboration is needed by security experts and AI researchers to create defenses against disinformation made by machines.
Last year, Choi took the helm of Project Alexandria, the “common sense AI” initiative backed by $125 million from Paul Allen, the late Microsoft co-founder and founder of AI2. She is also an assistant professor at UW’s Paul G. Allen School of Computer Science. In addition to Choi, the Grover team included researchers Rowan Zellers, Ari Holtzman, Hanna Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner.