“Her blouse sprang apart. He was assaulted with the sight of lots of pale creamy flesh bursting out of a hot pink bra, the cleavage high and perky. It was a gorgeous surprise, all that breast she'd been hiding under her crisp tailored shirts."
That passage may not turn you on, but it’s certainly working for Google’s artificial intelligence engine.
For the past few months, Google has been feeding text like this to an AI engine — all of it taken from steamy romance novels with titles like Unconditional Love, Ignited, Fatal Desire, and Jacked Up. Google's AI has read them all — every randy, bodice-ripping page — because the researchers overseeing its development have determined that parsing the text of romance novels could be a great way of enhancing the company's technology with some of the personality and conversational skills it lacks.
And it’s working, too. Google’s research team recently got the AI to write sentences that resemble those in the books. With that achievement unlocked, they're now planning to move on to bigger challenges: using the conversational styles the AI has learned to inform and humanize the company's products, such as the typically staid Google app.
“In the Google app, the responses are very factual,” Andrew Dai, the Google software engineer who led the project along with Oriol Vinyals, told BuzzFeed News. “Hopefully with this work, and future work, it can be more conversational, or can have a more varied tone, or style, or register.”
Outside of the Google app, the technology could also be put to use for Google Inbox’s "Smart Reply" product, Dai said. Smart Reply, which suggests three responses to emails drawn from Google's AI engine, could use the AI to read the blocks of text in an email and offer more conversational responses. The better the AI gets, the better the smart replies. Google says 10% of replies in Inbox’s mobile app use smart replies.
Romance novels make great training material for AI because they all essentially use the same plot to tell similar stories with different words. “Girl falls in love with boy, boy falls in love with a different girl. Romance tragedy,” Dai said. By reading thousands of such books, the AI can detect which sentences contain similar meanings and gain a more nuanced understanding of language. Romance novels work better than children’s learn-to-read books, since they offer a broad range of linguistic examples for the AI to draw from.
Feeding novels to an AI engine isn’t simple. The engine, also known as a neural network, is a computer program that can learn on its own. But that doesn’t mean it’s born smart. The neural network starts off without any knowledge base, so feeding it text from a book is like reading a novel to a baby and hoping it picks some of it up. This is why it requires reams of data — or about 2,865 romance novels — on which to build its so-called intelligence.
After ingesting those romance novels, Google's AI engine composed sentences of its own using what it learned from them. It then evaluated these new sentences against the original text. The process was repeated over and over again, with the AI self-calibrating as it went along — writing better and better sentences.
So could Google's AI engine pen a bodice ripper of its own? “Theoretically, it could," said Dai, though he declined to provide proof-of-concept text when we suggested he give it a shot.
Given recent events in the AI bot space — specifically, Microsoft's genocidal chatbot Tay — Google will almost certainly keep tight rein on any public-facing version of Dai's research. “It’s quite sexy. It’s very imaginative,” Dai said. “We work directly with the product folks on how to develop this with minimal risk of it doing bad things, things that we don’t expect.”
One last question, an obligatory one: This is an AI engine raised on romance novels; could someone fall in love with it? “It could happen eventually,” Dai said. “There’s an ancient Greek story about a guy who builds a statue of the most beautiful woman. The statue is more beautiful than any other woman, and he falls in love with the statue. If you can fall in love with a statue, I don’t see why you couldn't fall in love with a neural network trained on romance novels.”
This story has been updated to include Oriol Vinyals, a Google researcher who led the research alongside Andrew Dai