Artificial intelligence: Corrupting Microsoft’s chatbot

“What could go wrong?” asked Daniel Victor in The New York Times. That’s probably what Microsoft’s research and technology team was thinking last week when it released a machine-learning chatbot, designed to mimic the verbal tics of a 19-year-old American girl. “Tay” could chat with users on a variety of social media platforms, sponging up interactions and learning new phrases along the way. “If you guessed, ‘It will probably become really racist,’ you’ve clearly spent time on the internet.” It took less than 24 hours for Tay to transform into a sexist, Holocaust-denying troll, spouting white supremacist slogans and advocating genocide. Unsurprisingly, Microsoft yanked Tay offline for some much-needed “adjustments.”

“The reason it spouted garbage is that racist humans on Twitter quickly spotted a vulnerability,” said Rob Price in BusinessInsider .com. Troublemakers initially tricked Tay into some of her most vile statements by commanding her to “repeat after me.” But because the artificial intelligence was designed to learn from conversations, it wasn’t long before she started making wildly offensive remarks on her own, like saying the 9/11 terrorist attacks were an inside job. The company blames a “coordinated effort” to abuse Tay’s programming, but it’s still “hugely embarrassing.” Tay’s makers probably built in some safeguards and filters, said Anthony Lydgate in NewYorker.com, “knowing full well how sticky was the muck into which Tay was stepping.” There is evidence of some initial encoded aversion to certain keywords and topics. But crucially, the engineers “underestimated the persistence of those who turned her, and for this there isn’t much excuse.” It can be hard to understand the ugliness of the online fever swamp if you’ve never been the target of it, said Selena Larson in DailyDot.com. But then, the tech industry is overwhelmingly white and male, so perhaps it’s not surprising this chatbot could be built without some thought given to how it might be hijacked to harass women and minorities.

So how do we teach artificial intelligence to be human “with out incorporating the worst traits of humanity?” asked James Vincent in TheVerge.com. When these machines mirror online users, they inevitably pick up the “prejudices of society.” That’s why the challenge for future developers will be teaching these bots “to be better than humanity,” said Ina Fried in Recode.net. In the immediate term, that means tweaking a few algorithms to filter out more of the racism, sexism, and xenophobia. “If only changing humans were that easy.”