人工智能学会反驳偏执者

Social media platforms like Facebookuse a combination of artificial intelligence and human moderators to s�cout out and eliminate hate speech. But now researchers havedeveloped a new AI tool that wouldn't just scrub hate speech but would actuallycraft responses to it, like this: "The language used is highly offensive.All ethnicities and social groups deserve tolerance."

"And this type of intervention responsecan hopefully short-circuit the hate cycles that we often get in these types offorums."

Anna Bethke, a data scientist at Intel. Theidea, she says, is to fight hate speech with more speech—an approach advocatedby the ACLU and the U.N. High Commissioner for Human Rights.

So with her colleagues at U.C. Santa Barbara,Bethke got access to more than 5,000 conversations from the site Reddit andnearly 12,000 more from Gab—a social media site where many users banned byTwitter tend to resurface.

The researchers had real people craft sampleresponses to the hate speech in those Reddit and Gab conversations. Then theylet natural-language-processing algorithms learn from the real human responsesand craft their own, such as: "I don't think using words that aresexist in nature contributeto a productive conversation."

Which sounds pretty good. But the machinesalso s�pit out slightly head-scratching responses like this one:"This is not allowed and un time to treat people by their skincolor."
And when the scientists asked human reviewers toblindly choose between human responses and machine responses—well,most of the time, the humans won. The team published the results on the siteArxiv and will present them next month in Hong Kong at the Conferenceon Empirical Methods in Natural Language Processing.

Ultimately, Bethke says, the idea is tospark more conversation.

"And not just to have this discussionbetween a person and a bot but to start to elicitthe conversations within the communities themselves—between thepeople that might be being harmful and those they're potentially harming."

In other words, to bring back good ol' civildiscourse?

"Oh! I don't know if I'd go that far. Butit sort of sounds like that's what I just proposed,huh?"