Channel: Artificial Intelligence
Viewing all articles
Browse latest Browse all 1375

Here's how AI could fix the toxic nature of the internet



Venture outside the Safe For Work confines of the internet and you'll find off-the-scale offensive images that will rattle your brain.

Or you could provoke the ire of strangers on the internet, like Anita Sarkeesian, who runs a blog exposing the way sexist videogames depict and exclude women. On a daily basis, she receives sexual harassment and death threats from anonymous Twitter users. Last October, she had to cancel a Utah State University event because of an anonymous terror threat.

Because of incidents like these, Twitter and SRI International, the company that created Siri, are now deploying algorithms to crack down on online harassment and offensive material, according to Wired.

Online harassment covers a wide range of interactions, including personal attacks, offensive images, and threats on social media platforms. The most extreme forms of online harassment include stalking and doxing, when personal details like someone's home addresses are publicly revealed online.

According to a 2014 Pew Research Center Survey, 73% of those surveyed were the victims of some form of online harassment, with people ages 18 to 29 experiencing the brunt of it. Women aged between 18 and 24 were also overwhelming the target of online sexual harassment.

Currently, social networks like Facebook and Twitter require users to flag and report harassment. The problem is, the offending messages have then already taken their emotional toll.

In some other cases, content moderators are tasked with keeping offensive material from popping up in timelines and profiles. According to a 2014 Wired report by Adrian Chen, they "might very well comprise as much as half the total workforce for social media sites."

Shielding everyday users from offensive images and violent content is an emotionally and psychologically exhausting job. Content moderators, usually employed overseas in countries like the Philippines, spend 8 hours a day, day in and day out, staring at pornographic and gory images, Chen wrote.

With any luck, advances in artificial intelligence (AI) would make these content moderator jobs — and the internet at large — less toxic.

Norman Winarsky, SRI International's vice president president, told Wired that their AI could help filter harassment and offensive images out — much like the AI that's responsible for filtering spam in your email inbox — without the need for human moderators to see them.

"Social networks are overwhelmed with these kinds of problems, and human curators can't manage the load," Winarsky said.

SRI's AI systems build on the company's past expertise on how people communicate online using supervised machine learning. According to Wired, 250 people curated online harassment data from an anonymous social network and fed it into the AI system. Over time, the AI learned what words and phrases constitute harassment.

Eventually, SRI hopes to build a system that can even learn to flag sarcastic statements — which may seem harmless — as bullying by "analyzing patterns of actions," according to Wired.

While SRI's anti-cyberbullying AI may be available for practical use about six months to a year from now, Twitter is currently implementing AI systems to catch offensive images. This is especially important since images are shown natively in the network.

Last year, Twitter purchased MadBits, a company that builds image recognition AI systems, and set them on the herculean task of catching offensive images, according to Wired.

MadBits co-founder Clement Farabet told Wired that the system is pretty effective. When primed to catch 99% of offensive images, Farabet said it will "incorrectly flag perfectly acceptable pics just 7% of the time."

The AI system isn't Twitter's first line of defense, yet. Twitter's AI uses supervised machine learning to perfect it, requiring human employees to view and label offensive images before feeding the data into the system. The AI system is comprised of an artificial neural network, which mimics the interconnected structure of human brain cells.

But as the AI encounters more images, it gets better at distinguishing between what is and isn't offensive. Over time, "the need for this tagging diminishes," and fewer people will need to actually see the images.

SEE ALSO: One of Mark Zuckerberg's mind-blowing predictions about the future already exists

Join the conversation about this story »

NOW WATCH: Scientists have developed paint that changes color based on heat, light and impact

Viewing all articles
Browse latest Browse all 1375

Latest Images

Trending Articles

Latest Images