Nina (Zhuxiaona) Wei
On March 14, 2016, at 6:40 a.m., I woke up in the cold darkness and got washed and dressed. I took an Uber to the San Francisco Caltrain station and hopped on the 7:56 train, one minute before it left. The train arrived in Mountain View at 8:49. The first rays of sunshine were flying into my eyes, dancing around—“Welcome, Nina,” said the sun, smiling. “It’s like summer here!” I replied, smiling back. Another Uber drove me to the office. At 9:30 I stood up from the sofa in the lobby and stepped into the office…
“Welcome, Nina, to the AI Lab!” It was one of the most memorable moments in my life—I’ll never forget it. From that moment on, my journey of designing for AI products began. I felt lucky and grateful for being a so-called AI designer.
AI, artificial intelligence. One of the hottest and sexiest terms of 2016, 2017, and the years beyond. But this term makes me nervous—it makes me so nervous that I don’t want to say it aloud when I introduce myself to people:
So, what do you do?
I am a designer.
Oh really? Where do you work?
Hmm, I am working at an AI lab…
I say this word AI in a low voice, but you still hear it clearly.
AI?! Do you mean artificial intelligence? Cool!
This is what makes me nervous.
I am cool in a lot of ways, but not in the ways you’re thinking. I do not consider myself to be an AI designer. What exactly is an AI designer? “A designer who designs AI products,” you answer. What is an AI product? Isn’t it an artificial intelligence product? What is artificial intelligence?
The truth is that I am a normal designer. I design products you are familiar with, such as mobile and Web apps. My daily life is the same as yours—brainstorming, sketching, pushing pixels, or talking all day nonstop for a user study. Do you still think I am cool?
Even if my job is cool, I do not enjoy any privileges. Everyone is born to serve, either themselves or others. We designers are here to solve problems and make people’s lives better. The only difference between you and me is the problems we are solving. Most of the time, the problem is the same: There aren’t many truly new problems. We focus on one aspect of the problem and use different techniques to solve it, or a small piece of it.
What is AI? In my view, AI enables the augmentation of human intelligence. Products that utilize AI technology can be only AI-enabled products. A product is a solution for solving human problems. AI is not the solution; AI itself cannot be a product.
If AI were the core, we would become too attached to it. We would ignore the mission—solving human problems. Every technology is a tool but not the problem. As creators, we need to focus on the problem and detach ourselves from this fantastic tool. The tool cannot bring value until we find the right problems.
“AI is the new electricity,” says Andrew Ng, former chief scientist of Baidu. “Electricity transformed countless industries. AI will now do the same.”
There is a key difference between product thinking and design thinking: The former focuses more on the design of the problem while the latter focuses more on the design of the solution.
I first fell in love with AI several years ago when I was studying psychology in college. I was attracted by humans’ ability to develop artificial intelligence. But how can we bring the value of technologies such as AI to everyone? If normal people like you and me cannot enjoy the value of a technology in our daily lives, what is the technology for?
Products are one of the most important carriers of technology, but transforming technology into products is as hard as developing the technology itself. That is why I am here—to find the right problem and design an AI-enabled product in the right way.
Our product team’s mission is to innovate and incubate AI-enabled products. As the solo designer, I am responsible for all the design work as well as the product strategy. On March 14, 2016, the journey began. Four months later, on July 15, we released our first product, the TalkType voice keyboard (Figure 1). At the end of February 2017, TalkType downloads exceeded 100k. In the middle of March 2017, we released SwiftScribe, an AI-enabled transcription service (http://swiftscribe.ai/). Meanwhile, we are exploring and testing more product ideas. We tried and failed, again and again. The great feedback from users makes everything worthwhile. Lessons, tons of lessons, are being learned on this journey. Here are some of them.
Product thinking. Not everyone should be a product manager, but everyone needs product thinking. What is product thinking? “Think in products, not in features,” says Nikkel Blaase, product designer at XING .
Understanding the problem and user needs is the first step in the design process. User-centered design is user-centered: Users, not features, come first. Then how come designers think in features? I was confused. And what is the difference between product thinking and design thinking? I asked this question in Quora. Here is part of someone’s response:
Product thinking is a more holistic view than UX/design thinking…and it’s also much more complicated.— Joseph Dickerson, UX Lead at Microsoft .
I have increased my understanding along the journey. For one thing, Blaase comes from a traditional design background. Everyone is talking about UX, but much of the industry still operates in traditional ways: Product managers define products and write product requirement definitions while designers deliver designs, mostly the UI. Designers with UX Designer titles are actually UI designers.
There is a key difference between product thinking and design thinking: The former focuses more on the design of the problem while the latter focuses more on the design of the solution. When building a product, 80 percent of the time we are defining the problem. And we spend much less time designing the solution, as shown in Stanford DSchool’s design-process model (Figure 2).
Problem, user, and use case. We should not add technology to an existing product without clarifying the problem that this would better address. Take chatbots. Chat as a form of interaction is intuitive. MIT Technology Review acknowledged “conversational interfaces” as one of the top 10 technology breakthroughs . But that does not mean chatbots can be applied to every problem. Opera, Magic, Assist, Luka—so many bot startups were born in 2016 . Facebook, Telegram, Slack—everyone is building their own bot platform. Facebook has since made changes to its bot platform, with more emphasis on actions . It added a menu of features offered by bots, which allows for multiple and nested items to be built in. It is a much simpler experience without conversational capabilities. E-commerce sites are starting to pull their bots from the shelves, acknowledging that they didn’t do what they were supposed to . They are hard to use and do not provide personalized service.
Speech recognition can be the breakthrough: Bots cannot be disruptive unless we find the right problem. Bots are also not new. Joseph Weizenbaum from MIT built the first chatbot, Eliza, a computer therapist (https://en.wikipedia.org/wiki/ELIZA). Lacking natural language understanding/processing (NLU/NLP), Eliza is not considered to be a real conversational interface. However, with continual technological breakthroughs, I believe we will soon be able to have a real chat with a machine .
User. Who is our target user? Those cool young kids who love technology and must be hungry for new things? It turns out that’s wrong. From our data (user feedback and reviews), we noticed that there are a lot of not so young and even senior users. “Got this for my 92-year-old great grandfather and he loves it.” This review made my entire week. Users appreciate that TalkType solves their problems. For example, they might have shaky hands or fat fingers, or they are nearsighted.
“Speech to text” is the most searched keyword among all voice-keyboard-related searches. People who are looking for a solution must have a problem. They are the real users.
“I can type very fast with my eyes closed. Cuz I remembered every key!” said a young girl, holding her Samsung and showing us the default Samsung keyboard. Those cool young kids do not have a problem and they do not need another input tool. They are looking for cool new things. They are eager to express their unique personalities via the cool new things.
Use case. A use case defines the context for a problem as well as the relationship between the user and the problem. For example, note-taking has various use cases, such as class lectures, interviews, and jotting down random ideas. There is a balance between scoping the problem and scaling the solution. We can build a target solution for a specific use case, but since there are similarities among different use cases, we may lose the opportunity to scale up.
The following products (Figure 3) are solving a similar problem across different scenarios. They all use automatic speech recognition (ASR):
- Chatbaka Voice Messenger – the safer way to send messages while driving (http://chatbaka.com/)
- Cassette – transcribe, record, and share conversations (http://www.cassette.design/)
- Recorder – voice journaling made easy (https://itunes.apple.com/gb/app/recorder-voice-journaling/id1140324040?mt=8).
Test the value of AI. There are two broad contrasting approaches to build AI-enabled products. 1) Technology first: We have this technology. What kind of product can we apply it to? 2) Product first: We have this product idea. What kind of technology do we need? It costs more time and resources to develop new technology. We may already have technology that has not yet shown its value. So, we will start with the first approach to build a product enabled by the current technology. Then we will shift to the second approach to work with the research and engineering teams to build new technology to serve user needs.
There is a balance between scoping the problem and scaling the solution. We can build a target solution for a specific use case, but since there are similarities among different use cases, we may lose the opportunity to scale up.
But applying technology is not as simple as an add-on or a plug-in. The example products above all use ASR. Is the value of ASR the same across every product? Does ASR even bring value?
Adding AI to existing products can have a range of outcomes, including:
- Fails to solve the problem and hurts the user experience
- Fails to solve the problem
- Solves the problem
- Solves the problem and improves the user experience
- Solves the problem, improves the user experience, and brings extra value or creates new experiences.
We should avoid one and two. Three is not recommended. Four already can bring real value. Five is the goal—then AI is truly the new electricity!
Business value is as important as user value. But if we pay too much attention to business value, we may impose unwanted constraints. If our product can reach the level of bullets four and five, the business value will be clearer.
The two sides of AI. Microsoft Tay became a racist in less than a day; Google Photos appallingly identified two black people as gorillas. The more value AI brings, the more risks it has. Even a 1 percent error rate for an autonomous car can lead to deaths. At the other extreme, a smart light bulb may fail to turn on when someone arrives home; that person might then become a bit mad and yell at the bulb. But no big deal, right? The value it brings is small. Life continues.
The creator of Microsoft Tay ignored human nature, especially its bad side. There are two approaches to create Tay-like chatbots. The machine gets answers from rules based on user input. Or the machine will self-learn from the data. Tay uses the second approach. It is so smart that it learned racist language and became obnoxious in under 24 hours.
How about Google Photo? The creator may not have used a diverse enough dataset to train the model. Or they have not considered all predictable use cases and have not tested enough to uncover likely problems.
What can we learn from such accidents?
- Embrace human nature and design with all of humanity in mind.
- Testing, testing—always testing.
The simpler, the more complex. What is the best part of AI? It makes things simpler. Voice makes typing faster; search-engine algorithms help you to find whatever you want in seconds. Smart e-commerce recommendations know what you like.
How come it is so simple? Because it is complex. The simpler it is to a human, the more complex it is to the machine. Take voice typing. It looks simple—voice input and transcription output. It looks like the user does not have any interactions with the system except the input. In fact, the user and the system are interacting with each other all the time:
User turns on the mic… System starts listening… User starts speaking… System is gathering user’s input data… System starts transcribing… System finishes transcribing… System shows the transcription… User pauses for a moment… System is still listening… System may stop listening if there is no input for a while… User resumes speaking… System is listening… User corrects the transcription…
“Send.” The whole experience ends with this action. There are other use cases of course: What if the user speaks too fast? What if the environment is too noisy?
Think deeply and do not be fooled by the surface. Look into the experience of the user as well as that of the system. Since we are designing the interaction, we’d better pay attention to both sides to reveal the pain points and improve the whole experience.
Behavior change is hard. We can all be lazy. We tend to choose the default to avoid costs and risks (aka the “default effect”). Most users use the default keyboard that comes with their phones. “It is good enough.” “I never thought about looking for another keyboard.”
Behavior change is hard but it is not impossible. It should better serve users’ most urgent needs and guide them to a new and better experience.
Traditional keyboards provide an entrance for users to access the voice input. Unusually, voice input is the default on TalkType (Figure 1). We put the mic button up front. Users become more familiar with voice input thanks to its strong exposure. And there is a greater chance that users will tap the mic button and speak! Users also have ready access to the regular keyboard and shortcut keys such as numbers and emojis.
If I am alone, I may use it. But if it is in public, it is too weird.
What do you mean by “weird”?
Isn’t it weird to speak to your phone? And the phone will not speak back to you?
What’s odd about artificial intelligence is that it is artificial. It cannot be more natural than speaking with an actual person. But it is weird to speak to something artificial, though it should involve the same behavior. How can we make this odd experience more natural and user friendly? We need to find out the “why.” We need to know the user as well as the machine. We need to study not only the technology and the design, but also the sociology and the humanity.
Two of many researchers who have studied this relationship between humans and technology are Clifford Nass (https://en.wikipedia.org/wiki/Clifford_Nass) and Sherry Turkle (http://www.mit.edu/~sturkle/). I admire them immensely.
This relationship involves more than design, and far more than UI design.
You enrolled in several machine-learning courses on Coursera. You read every news article about AI. You bought an Amazon Echo and a Google Home. You want to be an AI designer. You ask me how.
First, ask a lot of whys: Figure out your inner desire. Why do you want to work on AI? Just because it sounds cool? Or do you believe it can change the world?
Just try: Try to use products that make good or bad use of AI; try to think about problems that AI can help to solve. And talk to people—AI experts, creators of AI-enabled products, users who like and are using those products in daily life. Then ask yourself a lot of whys again until you find the root cause of your want.
Learn smart, not hard: Knowing how it works is more important than learning how to do it. Our goal is not to be an AI researcher or engineer, but rather to design better AI-enabled products.
Collaboration is the key. It’s the beginning of the relationship between the human and the machine. There are a lot of challenges as well as opportunities. It requires AI researchers, software engineers, and product designers to collaborate and build AI-enabled products that can solve existing problems and create new experiences.
1. Blaase, N. Why product thinking is the next big thing in UX design. 2015; https://medium.com/@jaf_designer/why-product-thinking-is-the-next-big-thing-in-ux-design-ee7de959f3fe#.lg558qy4g
3. Knight, W. Powerful speech technology from China’s leading Internet company makes it much easier to use a smartphone. 2016; https://www.technologyreview.com/s/600766/10-breakthrough-technologies-2016-conversational-interfaces/
4. CB Insights. The rise of bots: A timeline of major VC-backed bot startups; https://www.cbinsights.com/blog/bot-startups-timeline/
5. Cohen, D. Here are the new features available to Facebook Messenger chat bot developers. 2017; http://www.adweek.com/digital/facebook-messenger-platform-1-4/
6. Pathak, S. Drop it like it’s bot: Brands have cooled on chatbots. 2017; https://digiday.com/marketing/brand-bot-backlash-begun/
Nina (Zhuxiaona) Wei is a product designer at an AI lab who has a great passion for products, psychology, and the relationship between humans and technology. firstname.lastname@example.org
Copyright held by author. Publication rights licensed to ACM.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2017 ACM, Inc.