5 Questions with Paulina Mustafa, Product Designer, Google

According to recent research, exponential improvements in and broader adoption of artificial intelligence is projected to more than double revenue to become a $12.5 billion industry by 2020. This represents a 20% annual growth rate putting AI in position to claim a total market cap of $120-180 billion.

Numbers aside, indeed AI is being used in more and more applications, however, ultimately it is a tool designed by humans. In this sense, AIs are like stories, constantly evolving and being shaped by the context in which they are developed, but we are still the authors.

On Friday, 1 November during #SMWLDN, Paulina Mustafa, Product Designer at Google, will discuss the design process behind AI, including why it’s never ending, and what we can do to proactively author the direction the industry takes.

We recently sat down with Paulina to learn about designing human-focused AI products, the AI experience as a craft, and much more.

SMW: What has been the biggest challenge in developing human-focused AI products?

PM: Humans and AIs process and express information incredibly differently. An AI is trained for a very specific task, and therefore can only process very specific things, albeit at extraordinary speeds humans aren’t capable of.

Humans can take in all kinds of inputs and make sense of them, even with imperfect information or ambiguity. We, humans, can do that because we intuitively understand the context around us— whether it’s cultural, social, or even physical context. Context is abstract, continually changing, and requires understanding so many bits of information. Humans can make sense of a new context upon encountering it for the first time, but an AI would need some kind of explicit training on how to recognize that pattern in order to make sense of it.

SMW: In your perspective, how can design help humanize AI experiences? What are some examples of design considerations that make AI more human?

PM: We’ve come to expect that machines are precise and predictable. When you put a complex multiplication into a calculator, you expect the correct answer, and you expect it every time. However, machine learning ‘ML,’ works a bit differently.

ML is based on complex probabilistic models, which means it provides the most likely answer, not the correct answer. The problems we’re solving with AI rarely have one single correct answer. Because there’s a degree of uncertainty involved with ML, it might make mistakes. To humanize AI experiences, we’ll have to create interfaces that allow for the AI to make mistakes and for the human to be unpredictable. We’ll need to consider ways of exchanging information between the AI and the human that allow for unexpected behavior in both directions.

SMW: AI is often talked about in the context of algorithms and processes that occur behind the scenes. Can you explain AI experience as a craft and why it’s important?

PM: While AI on its own is a collection of complex algorithms and processes, there’s so much more that goes into creating an application of AI. First and most importantly, humans need to decide what problem needs to be solved and whether AI is even needed. Then, humans need to create a clear definition of success, since all AIs are optimizing for something.

Finally, the AIs need to learn how to achieve that definition of success from a set of training data. That training data is critical because the model will derive its initial assumptions from it. The model is the only part of this that involves algorithms. All the other parts involve very careful and deliberate human decision making. I would absolutely consider that to be a craft.

SMW: What guardrails, if any, do you feel need to be placed on AI to ensure we are unleashing the full power of AI without becoming a detriment to human health and wellbeing?

PM: That’s a huge question! One we’ll be debating forever, or at least until AI is still relevant. Broadly speaking, we’ll need to consider policy and ethics, and we’ll need to develop both global standards that apply to everyone, and more local standards that allow for cultural differences to persist. More importantly, we’ll have to be willing to continually re-evaluate our policies.

Technology is changing at rates we’ve never seen before, and we’ll have to work incredibly hard to make sure our policy stays relevant and effective. Guardrails that are appropriate today may be much less relevant and less effective in a few years. Finally, I think it’s worth considering any kind of policy on the use of technology in general, not just on the use of AI.

One example of how we’re designing guardrails now is through the Google AI Principles. I talk about these themes daily at work. Some of the principles that I think are particularly relevant are that AI should be socially beneficial, that it should avoid creating or reinforcing unfair bias, and that it should be accountable to people by allowing for feedback and explanations. The principles also detail applications of AI we will not pursue, which include technologies that can harm people or be used as weapons. And as I mentioned earlier, the most important part about the principles is that they are dynamic and evolving, and should stay relevant over time.

SMW: AI experiences are increasingly embedded in our everyday products — many times without the user recognizing its AI. What are some key examples of this today? Do you believe companies and platforms have a fundamental responsibility to tell users when AI is being deployed (e.g. a chatbot experience that assumes the role of a human but is actually a bot?)

PM: Depending on how you define AI, an AI experience can be incredibly simple. It can be your new washing machine adjusting spin and time to dry your clothes well, or it can be your spam filter catching spam mail that you don’t want and hadn’t received before. There’s a running joke that’s maybe not a joke, which is that as soon as we use AI to solve a problem, it doesn’t seem like AI anymore!

Where I think it’s relevant and important to explain that AI is being used is when it’s an application of AI that is meant to mimic or replace human interaction, like in your chatbot example. From a user experience perspective, we believe that the better we can understand the system we are interacting with, the better the overall experience we’ll have.

One great example of this is the Live Relay feature that Google announced at IO. It helps people who want to make a call and, for a variety of reasons, may not be able to speak or hear. It takes the form of an assistive chat that can translate speech to text and text to speech. If you use this feature to make a call, it will identify itself as your assistive chat. That way, the person receiving the phone call will better understand how to communicate back, giving both all participants better outcomes and experiences.

SMW: One last bonus question. What is your favorite inspirational example of AI for good?

PM: I just read about this example and it made me incredibly happy. CSAIL at MIT and Massachusetts General Hospital created a deep learning model that can take a mammogram and predict whether that woman will develop breast cancer in the future. What makes it better than current guidelines and other algorithms is that it works just as well for black women and for white women. The beauty of it all is that the reason it works just as well on all skin colors is because it was trained using mammograms of women of all skin colors.

I find it so promising that AI can help us improve our medical system by allowing us to incorporate so much more demographic variety in the tools doctors use for evaluation. Link here.

Don’t miss your chance to explore evolution of artificial intelligence with Paulina at #SMWLDN (31 Oct – 1 Nov). Claim your pass by 9 August to take advantage of our early-bird rate before it expires.

Join 100,000+ fellow marketers who advance their skills and knowledge by subscribing to our weekly newsletter.

WATCH THE #SMWLDN 2019 PROMO