Five things to know about the next cycle of AI innovation - AI Creation

This week I had the opportunity to speak at the EmTech Digital Conference in San Francisco. I reflected on the state of AI today, and where we’re headed with its development. Let me share the five biggest things that I think we should be thinking about as we head toward the next cycle of AI innovation…

1)     Technology Innovation Is Cyclical—and Accelerating

I’ve learned over the years that technology innovation is cyclical. First, there’s an innovation. Then, there’s democratization. Then there is ‘realization’: What are people actually doing with this technology? Then, there’s responsibility: How do we best develop and deploy this technology responsibly?

Think about the car. It was first introduced in the late 19th century. It was democratized by Ford in the early 20th century. Legislation was passed that required seatbelts in all vehicles in 1968. But no state required you to wear one until New York made it a law in 1984!

The car brought benefits as well—freedom of travel for people everywhere and economic transformation. AI is much the same, with amazing potential, but accompanying risks. Today, we know some of the likely dangers of AI. We don’t have the luxury to realize what our responsibilities should be 80 years from now—we need to act today.

We’ve made amazing recent progress in AI, reaching human parity in industry benchmarks for speech, vision, and translation. These advances led very rapidly to implementation. But we didn’t always think through the societal implications, resulting in the issues we’ve seen with across the industry with facial recognition and bias, for example.  

At Microsoft, we’ve taken steps to address this issue. We’re working with customers globally and the industry to improve how this technology can recognize all faces. And in Microsoft Research, we’ve been working on something called synthetic data that can help to train machine learning algorithms with more complete and representative data to combat bias. You can read the paper here.   

As an industry, we have been working together to help correct many of these issues and implement better practices, principles and policy moving forward.

But by and large, we’ve been reactive. We must embrace the difficult questions not as an afterthought, but as a forethought. Because the next cycle is going to be big.

2)     AI Creation is the next cycle of AI innovation

The next step in AI’s development will be its ability to create, recognize human emotion, and generally behave more like humans do. I call this “AI Creation”—technology that can generate new media and engage in creative activities like storytelling.

Our social chatbot in China, Xiaoice, shows us what’s coming with AI Creation. Over the five years we’ve been developing her, she’s become quite famous. Since 2014, she has already engaged in over 30 billion conversations. And these conversations are lengthier and more meaningful than what we’ve seen in the past.

People really connect with Xiaoice because we’ve programmed her to pick up on the emotional side of conversation—what we call “EQ.” When we communicate, we use tone, wordplay and humor, things that are very difficult for computers to understand. But Xiaoice understands EQ, which makes communication much more natural and compelling.

Let me share a few more examples of how Xiaoice uses AI Creation.

Xiaoice writes and narrates children’s audiobooks, using deep learning to personalize everything from background music to story characters, based on input from families. Xiaoice can do this in just 20 seconds for a 10-minute customized story.  

Through machine learning, we introduced Xiaoice to the works of over 550 modern Chinese poets. She’s now writing her own poetry, and we’ve created a tool enabling anyone to compose a poem with Xiaoice’s help. If you give Xiaoice a picture, she’ll analyze the photo and suggest a poem based on the characteristics of the image. I’m a Computer Science Ph.D., and I’ve never written a poem in my life—until I did with Xiaoice.

It will take a long time, if ever, before AI Creation capabilities in creative fields are at the level of human output. But the real benefit is speed.

Xiaoice has machine reading and text creation capabilities that she uses to provide services for the financial sector. She can synthesize massive amounts of information to generate quarterly earnings report summaries for 90% of China’s financial institutions.

This used to take an analyst roughly 20 minutes. Now, Xiaoice can generate a high-quality draft for analysts in just a few seconds. And the analyst can edit and finalize the report in just a few minutes.

Xiaoice isn’t replacing people, but rather, augmenting what we can do, providing suggestions to make the act of creation faster and more productive, helping people to do more, faster. 

3)     AI Creation will bring new challenges…

As an industry, we’re essentially now at the stage of democratizing AI Creation. “The cars are on the road,” so to speak. This is where we could run into problems, as people use the tech in ways we haven’t anticipated or ways that could be harmful to society.

It’s already become so much easier for people to create fake voices, pictures, videos, or articles. But this isn’t new. People were “doctoring” photos long before Photoshop. Hollywood has used special effects to create illusions since the silent movies.

What’s different now is that AI can create a false “reality” so easily. And it is getting very good at it. Detecting fakes will require a very different set of forensic tools.

4)     …and we’re researching solutions

We’ve learned from the last AI cycle that we need to really consider our responsibility and societal consequences before driving forward. This is the point in the cycle where we need to engineer responsibility into the very fabric of the technology.

So at the same time we’re developing these technologies that can be creative, people are also working hard to develop technologies that can detect the fakes and misuses. The problem of detecting fake media is being worked on widely across academia, government, and business in the field of computer science.

There’s a lot of research happening to focus on how we can detect deepfake videos through things like lower resolution, blurs, or image disruption. Researchers are figuring out how to use deep learning and neural nets to better identify manipulations through these kinds of inconsistencies. In Microsoft research, our researchers are exploring ways to solve the tough challenges across images, videos, speech, and text.

5)     We must be proactive—and pre-emptive

We need to get ahead of some big issues that we know now will likely come to pass. We must be proactive. And in some cases, pre-emptive.

We need cooperation across industry, academia, and government. We need to educate consumers about where the content comes from that they are seeing and using. And we need to create a multi-disciplinary engineering culture that solves for these issues at the beginning of the development cycle.

Let’s not repeat the history of other technology cycles we’ve seen before, from the automobile to AI bias. We’ve got to move fast. We can’t wait 80 years for our AI Creation seatbelt.