Fluid XPS, one of our products, is an AI empowered digital shopping assistant. It aims to be the most knowledgeable sales assistant to make shopping recommendations by dialoging with shoppers, and relieve them from the paradox of choices.
So far, it’s great to see more than 50% users who left feedback claimed that XPS has found them an ideal The North Face jacket. But, we know there’s so much we need to improve and make it a trustworthy personal assistant to shoppers.
One of our technical challenges is to better understand natural language input & logics by users. At the moment, as natural language processing (NLP) is getting mature, and it remains a great machine learning training effort to our team, as well as many applications in this field. This can be a reason to compromise the “trustworthiness” in users’ minds.
The reason why NLP is not there yet is that people still haven’t been able to fully figure out and replicate how the brains work. Brains are brilliant inventions by the mother Nature, and this organoid is far more sophisticated than what a computer can accomplish: it’s amazing capabilities — how it learns, processes info and imagines — are still big mysteries to human beings.
That’s why when I heard that Aparna Chennapragada, Director of Google Now, was going to give an talk on building trustworthy AI-empowered products at the event Product That Count. I was definitely in.
From her experience of working with products that involve AI and lots of data, she categorized three dimensions on making such products trustworthy: AI (processing of data), UI (interactions), I (personalization for users).
For “AI” (processing of data):
1. If you solve problems that are hard for humans and easy for machines, it builds up trustworthiness.
For example, Google makes word corrections for wrong spellings, or provides you great results out of a world of information based on certain search keywords. An opposite example is making a robot use hand gestures or speak natural languages — humans do that much better than machines at the moment, while technologies on those topics are still far from ideal. However, people’s ambitions in this fields are going to constantly push it forward (we’re right there).
2. Study the “Wow/ What The Hell” ratio.
Keep an eye on how many users love and hate your data products and how the trend evolves. This gives you an idea how your product performances and whether you need to make adjustments.
3. Provide well-rounded data training sets for the product.
A lot of initial speech recognition systems only recognized male voices with US English, this quickly frustrated female users and users with accents. When you train your data products, make sure that you cover major user groups.
For “UI” (interactions):
1. UI needs to reflect how AI helps the user.
For example, when you Google some keyword with spelling error, it gives your spell correction suggestions “Did you mean…”. That’s when you know Google is working hard behind the scenes.
2. People may have different expectations for your AI products.
Some just want the outcomes and expect magics to happen, while others expect the results displayed together with explanations from the system, so they can tell how & why these results come. Find the balance between these two, from sources such as user studies, and try to get close to what your users expect.
3. Try to form a feedback loop.
Sometimes, user feedback is hard to capture, even though that’s what the product team definitely desires. For example, Google Now responds to user’s question “Will it rain this weekend?”. If this user finds it useful, there’s not a clear way on how she can give feedback on whether it’s useful or not. There’s no clear feedback loop back to the product team. Currently, there are supplement methods like user studies, but this is an area to be improved.
For “I” (personalization for users):
1. Make the personalization benefits clear to each of your users, because they will always look for “what’s in there for me”.
2. Allow users to teach the machine to learn.
Find ways to ask user if the content provided is interesting for them, or ask users what they are interested in. For example, after you upload pictures to FB, FB usually detects people’s faces for you, but sometimes, it asks who is it. This is an opportunity to teach the machine.