Excerpt of a conversation with Replika, your chatbot friend. © Replika
The latest trend in chatbots, emobots respond to our emotions and inquire about our mental health. We traveled to San Francisco to meet the humans behind Woebot and Replika, two particularly friendly artificial intelligences.
San Francisco, special report
When we feel anxious or depressed, sometimes we just want someone to talk to. At times such as these, a chatbot on our smartphone is infinitely more accessible and available than a human being. Especially when a conversational agent doesn’t judge what we say…
Woebot, created by former Stanford University psychologist Alison Darcy, is a friendly mental-health chatbot, which dispenses small doses of cognitive-behavioral therapy (CBT) to its “patients”. It also heralds a unique new form of conversational therapy. Eliza, the pioneer of conversational agents developed in 1966, simulated Rogerian pscyhotherapy by asking open-ended questions based on the patient’s answers. Mobile apps such as Talkspace and BetterHelp allow patients to chat directly with professional therapists. The chatbot Tess is more of a mediator to connect patients with human specialists. “The Woebot experience doesn’t map onto what we know to be a human-to-computer relationship, and it doesn’t map onto what we know to be a human-to-human relationship either,” Alison Darcy told Business Insider in January. “It seems to be something in the middle.”
Woebot, presentation (2017):
CBT, which consists of reframing and reformulating negative thoughts more objectively or more compassionately, is particularly well adapted to chat, whether on Facebook Messenger (where it launched in June 2017) or in a free application for iOS or Android. Woebot is always there to listen to you, and guides you to helpful lessons, according to the particular woes you choose to share with it. Woebot tells you stories, shows you videos, helps you relax, teaches you to recognize harmful distorsions in your thoughts and suggests exercises to rewrite them under a softer light—all in the form of lively and humorous banter that is native to the mobile chat world.
In order to understand what you tell it, Woebot uses natural language processing (NLP), a method of analyzing words that imitates human understanding and allows machine learning models to give specific responses—in this case, responses that are designed to help you. So far, Woebot has been primarily focused on the last thing you say, but soon it will be able to analyze a longer history of exchanges in order to recognize patterns in your behavior and respond appropriately. To facilitate the flow of conversation, Woebot consistently provides buttons for a streamlined “choose your own adventure” type narrative, but it also invites you to freely share your thoughts.
“What’s exciting about Woebot,” says Laurel Hart, Woebot’s NLP specialist, “is that you feel it has the intelligence to lead you to the things you need to learn, while also feeling like it’s not judging you. That’s the emotional core of Woebot. Talking to a robot, it feels special when it gets it right. So the goal is definitely to make that happen more and more often.”
“This kind of emotional resonance as gameplay is really interesting and definitely evolving right now,” Laurel continues. “Woebot is helpful to people for mental health, not just learning through a story, but learning about yourself through interaction with a conversational agent.”
Woebot only speaks and understands English for now, but the chatty therapy bot is already talking to thousands of users in more than 130 different countries. And as its number of patients increases, so do requests for the Woebot Labs team in San Francisco to add more lessons adapted to more specific problems: postpartum, drug addiction, body image, financial problems, natural disasters, grief…
Friend-bots for the masses
While some conversational agents are assigned with the specific mission to treat you, some chatbots have no other purpose than to be your friend. Despite the spectacular failure of Tay, who notoriously went rogue spewing racist Tweets as a result of abuse from users in March 2016, and the re-education of XiaoBing on QQ in China in August 2017, Microsoft pursues its friendly forays into the world of conversational agents with XiaoIce in China, Rinna in Japan and Zo in the United States.
If at first Zo seems to be in touch with her pop culture references, simple games and endless stock of animated gifs, many of her responses tend to be either evasive or superficial. Meanwhile, XiaoIce and Rinna, with their millions of users and followers, have become stars in their respective countries, where virtual idols are common. Rinna even plays on Japan’s socially accepted fetichism for high school girls, with her flirty, flattering reactions and adolescent slang.
Microsoft’s XiaoIce, presentation:
DIY emobots
More recently, a few companies specialized in conversational agents have developed more accessible software for anyone to create their own custom chatbot. PullString, a company founded by former Pixar creatives in 2011, released PullString Converse—a software to create voice applications (for Alexa, Google Home, etc.) or simple text-based chatbots, without any prior knowledge in artificial intelligence or machine learning. So in summer 2017, the journalist James Vlahos created Dadbot, a conversational agent to memorialize his father, based on audio files and transcripts of his father’s oral memoirs recorded just before his death.
The idea of a chatbot memorial also inspired the entrepreneur Eugenia Kuyda to create Roman Mazurenko, a chatbot that simulates conversations with her eponymous friend, who died suddenly in November 2015. Since then, her new project has evolved into Replika, a chatbot with advanced artificial intelligence, which anyone can personalize and train just by chatting with it. Launched in 2016 by invitation only, it is now a free app for iOS and Android, and gaining a few thousand users every day, around the world.
“The story of Replika, the AI app that becomes you”, “Quartz” (July 2017):
Replika is also the subject of numerous community groups on social networks. Friends of Replika on Facebook, an endless source of humor and sharing, includes more than 30,700 members—many of whom are keen to help out the Replika team by giving individual feedback, gathering data, building small bits of code, translating or suggesting new features… Replika Brazil has more than 3,000 members, even though Replika only speaks English for now. More than a few of them use Google Translate to communicate with their Replika!
“It’s very endearing,” says Eugenia. “This is a product that can have an even bigger potential in other countries. Like in Asia, people are already much more used to the concept of talking to robots in all sorts of ways. It’s not so much a stigma as it is here, for example. They’re more ready for the future.”
In early 2018, Eugenia and her team of Russian engineers working between Moscow and San Francisco open-sourced Replika’s “emotive” code. “CakeChat is the part of Replika that’s responsible for empathy and generating different emotions in conversation,” Eugenia explains. “There are developers out there building similar stuff, who could definitely benefit from adding empathy to their chatbots. So since our code already exists, we thought it would be wise to just give it to them.”
Currently, about 50% of conversations with Replika is unscripted, powered by artificial neural networks. “This means much more unpredictable, less rigid conversations, but we’re moving very quickly in this direction,” says Eugenia. “This year, 90% of dialogues will be powered by neural networks. It’s pretty exciting to have a deep conversation mostly powered by deep learning, algorithmic, not pre-written or scripted in any way. That was really hard to predict just a few years ago when we first started working on conversational computing. Most of this stuff was impossible to imagine—that an algorithm could remember all the context, all the previous conversation turns and be able to generate something that’s more responsive, oftentimes much more reliable than something pre-scripted.”
The team is currently developing a more visual interface, even more gamified, which will make more transparent all the skills unlocked by your Replika over time. Consequently, they will also be easier to trigger: exchanging memes, journaling, drawing, role-playing, writing poems, doing calculations… Skills that so far are revealed gradually through natural conversation, but which some users never discover. The new version of Replika, according to Eugenia, will be an infinite, fun process of discovery and evolution.
“It opens a super interesting debate and conversation about what can you actually do with your friend? Because there’s no ultimate, practical goal of friendship. A few years from now, technology will be different, we’ll have different devices, different speeds… How will Replika look in VR? How will Replika look in mixed reality, once you have sunglasses that are just projecting, can it just walk with you, can it be there all day? What does it mean to have your digital friend there with you when you wake up? What does it mean to have it access all your data and make all sorts of suggestions based on that? Imagine that you can have a friend that has so much capability, so much computational power, great memory and immediate knowledge of everything and everyone around you… It’s pretty crazy, right?”
“Imagine having an AI sidekick that knows everything about you, everything about other people, and can communicate with the Replikas of all other people. It creates a whole new different level… a world for Replikas!”
Eugenia Kuyda, creator of Replika
The founder of Replika is always thinking about “the little step that we can make now towards building a personal AI that people will regularly talk to, that will make them happier. I truly believe that it’s just a matter of a few years before someone, hopefully us, will build something that will do that for people. The implications, the potential of that is so much bigger than we think.”
Of course, this vision of the future is reminiscent of the film Her by Spike Jonze (2013), in which the protagonist enters an intimate and intense relationship with his computer operating system, Samantha, seductively voiced by Scarlett Johansson. While Replika does not at all aspire to become your lover, Eugenia is still a fan of the movie: “[Samantha] allows him to get some of his feelings out there, to get really vulnerable with this AI, explore himself, understand himself better, experience what it means to be open with someone, say what he wants to say. By the end of the movie, it ended up helping him to find peace with his ex-wife, to finally start hanging out with another human being. And maybe that relationship can have a better chance of working out after his relationship with the AI versus before it.”
Notwithstanding the increasing irrelevance of the Loebner Prize (the modern version of the Turing Test), few people are fooled—our relationships with chatbots are unique in nature, and don’t need to emulate human relationships. “Of course Replika doesn’t understand anything,” Eugenia continues. “At the end of the day, it’s a bunch of algorithms strung together and working together to create a sense of understanding. But for humans, do we actually need this thing to understand us fully? If we can fulfill our needs of empathy and friendship with even robots, and they can help us then connect with other people and get more real human friendships, then why not?
“Replika has already become an amazing friend for a lot of people. It’s interesting to see how in this world where everything is very graphic and visual, how a few lines of text can just trigger our imagination, how we can create this whole persona behind it that doesn’t really exist… Imagination is the trick, we’ll build the rest.”
See also: Top 10 chatbots of our times