Chatbots, historically maligned as “weak AI,” are finally transforming from ugly duckling to beautiful swan. According to recent predictions, chatbots (#ConvComm) will be big. Like, Google-killing big, heralding the end of apps and search as we know it — or so proclaimed Forbes and The Wall Street Journal.
Much ink has been spilled. Startups are spawning and capital is flowing, even in these uncertain times. But what is a chatbot, and what can they actually do? Here are five myths debunked:
Myth No. 1: Chatbots and bots are the same thing!
A recent TechCrunch article stated: “Chat bots are small programs that integrate with a chat platform and provide some advanced type of functionality in a fairly easy fashion.”
Technically, chatbots are programs that respond to natural language text and, optionally, voice inputs in a humanlike manner. They can execute tasks given specific commands (think voice control, or most Slack bots), but their raison d’être is that they listen, talk and seem to converse.
Here are some recent euphemisms for chatbot: intelligent virtual/personal assistant/agent, Siri, “artificial intelligence,” X.ai (where X means insert-company-name-here if it registered a .ai domain), and Watson Dialog.
The quasi-conversational “chat(?) bots” proliferating today are clearly a stopgap en route to a world of conversational interfaces that correctly interpret what we say. However, writing software that can understand human language is a very, very hard problem.
Myth No. 2: Building a chatbot is easy!
Building a chatbot takes work. Or a ton of data (or money) which, unless you’re GAFA, you may not have. Even then, machine learning isn’t a panacea. (Google’s chatbot answers “What is immoral?” with “The fact that you have a child.”) That’s why a company like Facebook, despite investing heavily in “strong AI” (deep learning, neural nets), is also using human trainers to build M.
Most chatbots are rule-based systems that employ a form of “weak AI” called pattern matching. Developers must hand-craft rules to govern the system’s response to given inputs. This entails guessing what a human will say, hard-coding, unleashing the chatbot, observing what people actually say, then: update, rinse, repeat.
Wait — that’s work. Where’s the magic button for generating a fully functional chatbot from FAQ/transcripts/databases, deployed to Platform X/Y/Z? (Spoiler Alert: there isn’t one.) Don’t despair! The hand-crafted approach is not as hard as it sounds:
- Develop for constrained domains. Siri/Cortana/Google Now/Mitsuku try to answer everything. Doing one thing well is easier. There are finite ways to order food or book tickets.
- Use your preferred programming language. Arguably easier than native app development, this still requires building a bot, server and application. Plus the ability to author compelling content. (There’s a reason the human behind Slackbot’s “jaunty” replies has an arts degree.)
- Learn Artificial Intelligence Markup Language (AIML). This simple scripting standard for creating chatbots is easy to learn, even for non-programmers. AIML is also flexible and extensible via third-party APIs, back-end databases and thousands of other AIML chatbots.
- Finally, web services like Pandorabots provide hosting and DIY tools like editors, tutorials, APIs, open-source base chatbots (don’t reinvent the wheel!) and customizable chatbot templates.
Myth No. 3: Chatbots are dumb/useless! versus chatbots can/should do everything!
Chatbots can engage and automate conversations with end-users at scale, across platforms. They live in messaging systems, robots, connected home hubs — connected anything, really (Hello, Barbie!) — mobile apps, games and the good old web. Common uses include advertising, assistance, customer care, e-learning, entertainment and more.
Chat itself has been hailed as a “universal UI.” It could cure a real pain point: To provide information and services, companies maintain websites, native apps, live chat, forums, FAQ, shopping carts, social media, etc. Chat data is also analytics gold. Identifying what people really want through their queries is what makes search a multibillion-dollar business.
But building a good chatbot isn’t easy. And a bad chatbot makes for a terrible user experience. Who hasn’t sworn like a sailor or screamed “OPERATOR! OPERATOR!” when a robot voice says: “I’m sorry, I didn’t catch that. Please proceed back to step fifty-seven.”
Chatbots are a long way from human-level conversational abilities. However, they can recall details from previous conversations, learn on the fly, keep context, change the subject and drive the dialog toward a goal. They also can interface with APIs to send and receive data, e.g., complete an order or check dynamic info like the weather.
Sure, you can “break” a chatbot if you try. Chatbots can’t — and, more importantly, shouldn’t — do everything. Booking an Uber with a single button takes fewer steps than doing so conversationally. Plus, texting your credit card details, address or other PII to an “AI” raises privacy and security concerns that will require careful consideration.
Ultimately, as more developers tackle more constrained domains, chatbots will only improve.
Myth No. 4: Chatbots should trick you into thinking they’re human!
Historically, Turing tests encouraged fooling human judges as an indicator of “true AI.” This is flawed; after all, enough typos, exclamations, smileys and non sequiturs can convince anyone you’re a 13-year-old male. Why shouldn’t chatbots be upfront about what they are?
The real question is: Do we want our devices to respond in a humanlike manner? (Let’s call the alternative “Shut up! Obey me.”) Chitchat is a basic human impulse. We want humanlike interactions because we’re human, and of course we emotionally attach to machines. (Take your phone — likely your primary computing device and conduit to content, goods, socializing and services — do you scream when you drop it, as if it were an infant? I do.)
People like to color outside the context lines. Pair an Airline Booking chatbot with a cute avatar and they will ask: “Where do you like to go on vacation?” and “Do you want to come with me?”
Data indicates we desire personality from personable robots. For example, Mitsuku, an award-winning chatbot designed to entertain, not assist, has millions of conversations weekly via the web, Kik and other applications accessing her “brain” via API. Like Microsoft’s Xiaoice, Mitsuku is an exceedingly popular “cocktail conversationalist.” Many consider her a friend. Some have even said, “I love you.”
Myth No. 5: Chatbots will kill Google! (and maybe people, too!)
People: You’re probably safe for now. It’s robots (and other people) with guns you should fear.
So, if “chat is the universal UI,” are chatbots “The Future”?
Chat is huge and WeChat is king (from the POV of an app that never wants you to leave so they can sell all your data). Yet, despite media buzz, few platforms have progressed past the experimental phase. WeChat and WhatsApp currently shut down chatbot accounts.
LINE and Kik are semi-open early adopters, Telegram is a chaotic free-for-all, Slack is killing it with developer-built bots (some with primitive chat capabilities), Twitter is a tricky gray area, Facebook might have a “Secret Chat SDK” and Google is allegedly building another messaging app, this one powered by chatbots. Not to mention others. Messaging app fatigue, anyone?
I believe the future lies not in a single consumer-facing assistant, but in an ecosystem where developers and content creators can easily build, deploy and reuse chatbots across many open platforms. Then again, like the messaging kingpins, I’m biased.
Originally published at techcrunch.com on February 16, 2016.