The capabilities of Google's artificial intelligence are staggering. The phantoms within Google Photos can organize your pictures, the wizards inside Google Docs let you type and edit using spoken commands, and the brilliant pixies powering AlphaGo can easily beat a master of a 2,500-year-old game more complex than chess.
Now it's turning that immense power loose in the browser-based version of Inbox. Whether you're using Google's sleek email manager in the browser or via mobile app, Inbox can now answer your emails for you. The machine replies aren't sent automatically, which is good. Instead, you choose from three responses suggested by Google's AI. (It works in some messages, but not all of them.) You also can use the suggested responses as starting points, editing or adding text as you like. The browser version of Smart Reply rolled out this week, about four months after it appeared in Inbox for Android and iOS.
Google's neural networks, which are programmed to learn behaviors through training, do the work. In my hands-on tests, Smart Reply could tell the difference between an email from a real person and a spam blast, offering one-click responses for the former but not the latter.
"That first decision is made by an artificial neural network very much like the ones we use for spam classification and separating promotional emails from personal ones," says Greg Corrado, senior research scientist on the Google Brain Team. "Our network has been trained to predict whether this is an email someone might write a brief reply to."
From there, Smart Reply generates a trio of one-click responses in the "Reply" field. The company says its engineers don't have access to the actual text of your messages; Smart Reply scans a message that hits your inbox, decides whether it's worth answering, and formulates three responses with its neural network.
Corrado says the implications of the technology will eventually be far greater than keeping your inbox in check.
"Part of the reason machine intelligence is so exciting right now is because the ideas and principles apply very generally, even when the detailed system does not," Corrado says. "The research ideas behind Smart Reply were originally conceived of to improve machine translation, not to power an automated email responder."
All the responses I encountered were impressively human-sounding and on-topic, but Smart Reply has limits. Most Smart Replies are appropriate for yes/no questions, while open-ended queries can churn out hilariously passive-aggressive results. The question "What do you want to eat?" generated three response options: "Whatever you want." "What do you want?" and "I don't know." It's like dating a goth.
The consumer version isn't tuned to parse answers to an A or B question, but Corrado says an internal version can handle it. For example, the question "Do you prefer celery or artichokes in your salad?" spat out the non-committal responses "Either one is fine," "Both sound good," and "Whatever is best for you." It did not offer the option of choosing celery or artichokes. Still, I like both, so maybe the artificial intelligence is onto something here.
"The 'research grade' version of Smart Reply, which is allowed to say anything it wants, already gets things like this correct much of the time," says Corrado. "Unfortunately, that system also says a bunch of other things that, while hysterically funny on occasion, are not consistently useful enough to put in front of millions of users."
Some open-ended questions bring up spot-on responses. Although Smart Reply doesn't cross-check your calendar and make decisions, it punts some decisions back to the message sender.
"It's not by design exactly, but instead largely a consequence of restricting Smart Reply to generate short replies that fit in the UI," Corrado explains. "Phrases that kick the question back to the other party are often short, and therefore well represented."
For example, "Can you call me tomorrow?" brought up responses of "Sure, what time?" "Sure, what's up?" and "I won't be available tomorrow." A question about rescheduling a meeting suggested "How about next Tuesday?" Tuesday is free in my calendar, which led me to believe Inbox was actually looking at my schedule. But that was just blind luck: The other options of "I am available any day next week" and "I'm available anytime" weren't true.
Although he wouldn't reveal any plans to build a calendar-checking part of the system, Corrado says it's one of the things he wants Smart Reply to do.
Smart Reply shines most brightly in simple yes/no queries. When I received an email asking whether I was rap-squatting in front of a bicycle, Inbox ginned up three enthusiastic responses: "No, I'm not!" "Yes!" and "Yes, indeed!" It doesn't just stick to the same three responses for the same question, either. A follow-up email that repeated the rap-squatting question generated different response options: "No, I'm not!" "Of course I am!" and the emphatically uncertain "I think so!"
Potential responses are meant to get better over time by learning from the replies users select, and there are contextual cues that influence the tone of each Smart Reply. Corrado says the formal or informal tone of each reply is based on that of the incoming message. That sometimes involves adding an exclamation point to informal responses, although those decisions aren't based on a single person's writing style.
"Smart Reply doesn't tie information to individual contacts or learn about individual behavior," Corrado says. "Style matching is an active area of artificial intelligence research, but personalization at the level of an individual user's style isn't something we have incorporated into Smart Reply---at least not yet."
Google says 10 percent of Inbox responses on the mobile app are driven by Smart Reply. Depending on your perspective, that's a testament to how well the technology works or a damning example of how impersonal we've all become. Corrado says the intent behind Smart Reply is to save time "replying to simple logistical emails" rather than to replace thoughtful, longer messages. He hopes it also helps its users spend more time in the real world.
"The less time I spend typing 'See you soon' with my thumbs on my phone, the more time I have to actually enjoy the walk down the street to meet my friend," Corrado says.
