Mobile chat is where it’s okay to be dumb. Your LinkedIn page may be a work of art wrought in HR-friendly prose; your tweets, so sharp you could shave with them. But texting? Let the autocorrect errors fly! Why punctuate? Isn’t that what old people do? In fact, why use words at all when any number of pictographic pokes, nods, and grunts available at the press of a button will get the job done faster? Blame the medium, says Jason Cornwell, the chief user experience designer behind Google’s communications apps. “Chat is inherently limited and low-bandwidth,” he says. It’s also here to stay: “Our phones are basically chat machines at this point. That’s the dominant activity that virtually everyone does on their phone.”
Allo, a mobile messaging app launched by Google in 2016, is the company’s attempt to use machine learning to make chat, if not smarter, then at least a hell of a lot more useful and expressive. Machine learning and artificial intelligence are becoming the engines behind nearly everything at Google, from Gmail’s spam filters to AlphaGo, the neural-network-powered software that recently beat the world’s best player at Go (a 2,500-year-old strategy game long considered impregnable by AI). As part of CEO Sundar Pichai’s strategy of transforming Google into an “AI first” company, Cornwell’s design team was charged with building “an app that was about chat on your phone, but at its core was about machine learning,” Cornwell explains. Exactly what machine-learning technology can do for people pecking out slang-filled, emoji-studded missives to one another is now his job to figure out.
Cornwell had navigated a similar design challenge two years earlier as user experience lead for Gmail’s offshoot app, Inbox, which uses machine learning to power its Smart Reply feature—those little rectangles of suggested text that you can select instead of manually typing your own message. Bringing Smart Reply technology to chat via Allo seemed like an obvious extension, because “editing text on a phone is still painful,” Cornwell says. Existing predictive-text functions and autocomplete help you type individual words slightly faster, but you still have to do the composing. Allo’s approach aimed to leapfrog that step entirely. “The goal was to think of it as a smarter extension of autocomplete, to help you say the thing that you already were thinking about saying, as close to in your voice as possible,” he explains.
That’s where Google’s machine-learning capabilities come in: The more you use Allo, the more its algorithms can ascertain what you sound like and generate prewritten responses that don’t sound canned. What’s more, Allo can learn how you text with different recipients—so it can offer up a “nice dude” in response to your best friend, but not when you’re messaging your mom. You can’t edit Allo’s Smart Replies, though, so you’re stuck with using—or ignoring—whatever it serves up. But that’s on purpose: Chat is “a rapid-fire medium,” says Cornwell, and testing showed that “it’s almost just as fast to type something new out” and send it on the heels of a Smart Reply that isn’t quite perfect.
Smart Reply also has a stealthy purpose: to introduce users to Google Assistant, the real brains inside Allo, along with the company’s new Google Home smart speaker. If you’ve ever been forced to pop out of your messaging app in order to Google something—say, the location of the restaurant where you’ll be meeting friends, or flight prices for a vacation you’re planning with a loved one—you’ll understand the assistant’s appeal. In Allo, you can just text your query to Google in natural language, as if it’s another person in the chat thread. This user experience is authentically conversational—the assistant doesn’t use punctuation in its texts, either!—and it’s so seamless that newcomers might not even realize they’re using this separate and sophisticated tech product. (To help encourage users to experiment and discover further capabilities, Smart Reply provides a button to tap that invokes the assistant in an Allo chat window.) It also avoids the dreaded “Microsoft Clippy” problem: Instead of a chirping robot interrupting your conversation to be “helpful,” Smart Reply creates opportunities for the assistant to introduce itself organically. For instance, when a friend messages you asking if you’d like to grab a bite downtown, Smart Reply may offer up a “sure,” “nah,” and an option from the assistant suggesting a search for local restaurants. “It’s not like the assistant is jumping up and down at you,” Cornwell says. “It’s offering you, in this native format, the ability to take the next step.”
The discreet way Allo encourages interactions with Google Assistant informed later designs on the assistant’s stand-alone app (available on both Android and iOS). “When you chat with the assistant in Allo, we also prompt you to ask the next question—so if you text ‘weather in mountain view,’ the assistant would provide the weather for that day and then offer up more specific phrases you could tap on, like ‘how about this weekend?’ ” Cornwell explains. “We found that when people were first using the assistant, they would structure their questions very specifically to get the exact answer they wanted, similar to how you’d type a query in Google search. But we wanted to help people learn not only the types of questions they could ask, but also [how to] speak more conversationally with their assistant.” The more casual a relationship a user has with her assistant, the more likely she will be to interact with assistant-augmented products that don’t rely on typing or screens at all, like Google Home.
But Allo’s most imaginative fusion of chat with machine learning, which it rolled out as an additional Allo feature in 2017, isn’t about helping you respond faster or getting things done. It’s about helping you put your best face forward. Just snap a pic of yourself in Allo, and within seconds Google’s machine-learning capabilities transform it into a suite of 24 “selfie stickers”: cute cartoon likenesses (created by Lamar Abrams, a storyboard artist for the Cartoon Network) that you can text and share like hyper-personalized emoji. “There’s a component of chat that’s about identity and how you see yourself—the craft and care that you put into your own communication,” Cornwell says.
Bitmoji—a popular third-party app that integrates with Snapchat—lets users create similar avatars too, but only manually. Allo’s automated version gets arguably as close to a good likeness as Bitmoji’s does—but not too close. And that’s by design, according to Cornwell. “Even if the algorithm was perfect, people wouldn’t feel good about [selfie stickers] unless they could put their personal stamp on it,” he explains. In other words, there’s something about tweaking a caricature of yourself that’s utterly essential to trusting it. Cornwell draws an analogy to instant cake mixes from the 1950s, which required that you crack an egg into the bowl: “There really is no technical reason to add that to an instant cake mix, but it just didn’t feel like cooking unless you broke an egg. In machine learning,” he explains, “there’s this key question that designers need to answer: Which pieces do we automate for you, and which pieces do you need to do yourself in order to feel good about the end result?”
Indeed, Allo provides enough combinations (about 563 quadrillion, if you want to get specific) that you could literally customize your selfie sticker until the sun burns out and still have plenty left to try. But really, what matters most to people “is the hair,” Cornwell says. “It’s one of the most critical features in terms of making people feel like they were seeing an authentic version of themselves on-screen,” or at least the version they wanted others to see. It’s true: When I tested a selfie sticker, my machine-learning-generated cartoon hair looked better than the real thing. As Cornwell and his team continue to expand Allo’s capabilities—the app has only a fraction of the user base that iMessage and Facebook Messenger boast—he understands that, when it comes to texting, it’s not just what you type, but how you look.