Designing Human-Centered AI Products  (Google I/O’19)

Designing Human-Centered AI Products (Google I/O’19)


[MUSIC PLAYING] KRISTIE FISHER: Hi, everyone. Thank you so much for
coming to this session on designing
human-centered AI products. My name is Kristie
Fisher, and I’m a UX researcher on Google Ads. And I’m also a contributor to
the people in AI research team. JESS HOLBROOK: Thanks. And I’m Kess Holbrook. I’m one of the leads of our
People + AI Research team here at Google. We usually go by
the acronym PAIR. So I’ll be calling
us that from now on. And the PAIR team looks at
human and AI interactions across a number of domains. And there’s something that
we’ve noticed recently that I’m guessing you have as well. AI is everywhere right now. AI is everywhere. It’s in the products
that we use every day. It’s at the forefront of
just about innovations in just about every
industry we can think of. And obviously, we’re seeing
more interest in AI development from developers like yourselves
than we ever have before. But beyond that,
if you’re at I/O, we’re guessing that you want to
accomplish a couple of things. You probably are
interested in AI if you’re in our
session of course. But we’re guessing that
you want to do something big and important with it. And we’re guessing
that you actually want to put the user first
at every step of the way. At the same time,
every developer wants to manage their time
and resources to make sure that they’re building something
that people are going to love, that works great, and that you
can build a business around. And so today, we’re going
to talk about a resource that we launched today to help
you do both of those things. We’re calling it the People
+ AI Guidebook, which is live today, ready to go, and
something we’ve been working on for about the past year. The guidebook is
a toolkit that you can use to ensure that the time
and energy and resources you put into development can make– AI development– can
have a big impact and put the user
first the entire time. It’s the result of contributions
from over 100 Googlers and is really a
synthesis of everything that we’ve learned
about how to practice human-centered approach to
AI over the past few years. The guidebook sits already
in an ecosystem of resources. You know, we’re not the
first piece of guidance around AI that Google
has provided to you. So the best way to think
about the guidebook is we– on one side, we
have the AI principles. These are our North Star. These are what guide every
major decision we have. On the other side, we have
the basic frameworks and tools that you already know– TensorFlow ML Kit,
all these things. The guidebook is really a
connection between the two. It’s a way to actually put
a lot of the AI principles into practice in a different way
of approaching your work when using the tools
that are out there. It’s a way of working versus
another technical guide or design spec or
something like that. The guidebook has
six chapters, which really cover the entirety
of the development process from identifying and defining
user needs and success, to how to collect data to build
the right kind of AI model, how to explain your AI to users
once it’s been developed, all the way to the inevitable
failure of the AI, and how to do that
in a graceful way where people don’t
abandon your product and still want to
continue to use it. One of the things we’re
really proud of as well is the guidebook content is
very tactical and hands on. You should be able to go back– I feel like you should be
able to read it in the morning and go back to your
desk in the afternoon and put it into action. It’s chocked full of
worksheets and things that you can print out and use
with yourself and your team as well. There’s more in the
guide book than we could cover in one talk
or many, many talks. So today, we’re
really just going to focus on three chapters and
one insight from each chapter that we think are some of
the most important ones to think about if you want
to be practicing developing AI in a human-centered way. First, we’re going to pull from
the user needs and defining success chapter and talk about
how to identify if AI even adds unique value to your product. Despite what you
might hear, you don’t have to add AI to every single
thing that you’re building. Second from the data collection
and evaluation chapter, we’re going to talk about how
to translate the user needs that you identify into data
needs, finding that data, and training the right model. And from the explainability
and trust chapter, we’re going to talk about ways
that you can explain your AI and what you’re AI is
doing to users in ways that they will understand and
feel in control while using your product. So to start, we’re
going to talk about how to identify if AI adds
unique value to your product. And to do so, we have to walk
through a few big questions you have to ask yourselves. So first, is, is there a user
need and associate pattern of behavior with that need? Next is, is there a definition
of success to optimize for? In this one, we
know that there’s lots of metrics you could
optimize in the AI for. We’re going to talk
about how you do that in a people-centered way. After that, we’re going to
look at how you think about, given the scope of the
need you’ve identified, could a non-AI solution work. Like I said, we don’t have
to AI every single thing. And lastly, if AI is
needed, which type is best? AI exists along many
different spectrums. One of them is really
from almost automation to augmentation. We’ll kind of talk
about those two ends. A lot of this is
shifting the mentality from a technology-centered
one to a people-centered one. One of the ways we
think about this is right now a lot of people
approach a problem and say, can we use AI to blank, whatever
they’re thinking their business problem is. And it’s really moving to how
might we solve blank or solve for blank, and then can AI solve
this problem in a unique way? That’s moving things too much
more people-centered version of AI. So everything I’ve
said so far has still been a little abstract. And one of the ways that
we hope that we can really communicate this clearly to
you is through an example. So just to kind of
start us off and get everybody a little loose, who
here flew to get here to I/O? OK, quite a few people. Kristie– yeah, me too. And of the people who
flew, how many of you like did some
research to figure out what was a good price for
airfare before you came here? OK, and then lastly, did anybody
use Google Flights to do that? OK, a few. Thank you for the business. We appreciate it. [LAUGHTER] So Google Flights team– the Google Flights
app has a feature called flights insights. And so it’s a
feature that uses AI to help you know when the
best time to buy a flight is. And this team
actually used insights from the guidebook in
the guidebook itself over the past year to
build this feature. So throughout this
talk, we’re going to introduce some guidance
from the guidebook but then use the Google Flights
case to really ground it out and let you know
how another team already used this, this content. OK, getting back to the four
questions we just had there. So first, you need
to ask yourself, is there a clear need and
associate behavior pattern? If you’re at all familiar
with human-centered design, it always starts with– starting with a user need or a
human need or a person’s need. Starting with there is
basically guaranteeing that you’re going
to find something interesting along the way. It may not be that exact thing
you find at the beginning, but you will get there. If you have the ability to
do formal user research, we of course encourage that. If you don’t have that ability,
or don’t have the resources, just talking to some
of your current users or prospective users
will really help you understand their needs. Then there’s the associate
behavior pattern. One of the really
interesting things there is it gives you a baseline
for the state of the world. But it also starts to give
you the scent of the trail that you want to be on
to find the right data that you’ll need later
to train your AI model. In the flights
example, they were– if we go the flights example,
they had a really clear need. They knew from
their user research and talking to their
customers that people wanted to know when to book a
flight for the best price and not get burned. They also had a really
good knowledgeable– they already knew the
pattern of behavior people were using at the time. So people would go back to the
same sites over and over again at different times
because people were trying to figure out
when is the best time to buy? Is it Tuesday night? Is it Saturday morning? Is it a full moon? I don’t know. I can’t figure this out. And so they had some data
there, and they had some needs that they knew that
they could address. Next, they moved on to, is there
a clear definition of success to optimize for? So we all metrics like
views and some engagement and clicks and things like that. Defining success in
a people-centered way is a lot harder than
you might think. We spend a good amount
of time about this on this in the guidebook. You want to be focusing
on some harder to measure but more important things like
are you increasing your users well-being? Are you increasing their long
term success and long term engagement with your product? And things like are you
making them happier? Are they enjoying
your product so much they’re going to go
tell somebody about it, which is a lot harder. You also want to make sure– sorry, behind one. You also want to make
sure and watch out for secondary effects. So some of the
secondary effects are things like what if you’re
too good at the metric you optimize for? Or what if it’s done
in a malicious way? So one of the ways
that this happens is like click bait articles. Like they have really
good engagement, and they have all these numbers. But they actually
disappoint a lot of people and are negative for
the overall ecosystem. And we get into this a
lot in the guidebook. Another way you can think
about defining success is through the good
old confusion matrix when considering any
kind of AI model. I’m guessing we’re not going
to get into this in any depth here, but this kind
of underlies a lot of the predictive
AI models where you have the reference,
which is the true state of the world, either something
is or isn’t in a category. And you have the AI prediction
that it is or isn’t. And so there we have our true
and false, true positives and negatives, and our false
positives and negatives. In a typical model, those
are considered equal. But when you’re thinking
about the user experience, you may actually want
your model to focus on one versus the other. So if we move back to the
Google Flights example, this was their confusion
matrix for flights prediction. So they’re trying to predict
is the price of a flight going to go up in the future? So we have our two outcomes
here that the price, in fact, does go up in the future. And the AI model
thought it would. And that the price does
not go up in the future, and the AI model
thought it would not. All good. No problems there. When we get to the errors,
though, something interesting happens. So in the case where the model
predicted the price would go up but it didn’t, that’s actually
not that bad of an error because the person would
just buy the ticket at the price at the best price
they could but maybe just a little earlier than they
had to, not all that bad. The really bad one is when
there was a price increase and the model decided– and the model did not
think there was one, right? Because this could cost
somebody a lot of money or even miss their flight
or something like that. We don’t just have guidance. We also have a lot of
tools to help you actually inspect the data. One of those is something
from the PAIR team that’s called the What If tool. So the What If tool
allows you to upload and dig into your
data and explore the consequences of false
positives and negatives, allows you to examine different
precision and recall thresholds and associated trade offs,
and even compare the output from multiple models. So you can start to
understand the errors an AI model is making
for whom and in what context a lot better. So going to the flights
example, is there a clear definition of success
to optimize your AI for? Success, yes, buying at the
right time at the right price was the clear definition
of success for them. Was there something
they could optimize for? Yes, the best time to buy– and like we
mentioned, it was all about avoiding those false
negatives or those cases where the model did not think the
price would go up and it did. Next, we want to consider,
given the scope of the need you’ve identified, could
the non-AI solution work? The Google flight team
had a Non-AI solution working and in place, but it
was extremely hard to maintain. And the accuracy was hard to
keep at users expectations. They also have millions of daily
active users and cache flights up to a year in advance. So there’s a huge complexity
to this to their need. So the scope is appropriate
AI is needed because there’s just too much complexity to
scale their current solution to their users needs. But still it was really grounded
in that maintenance of meeting their users expectations and
the engineering sustainability of a heuristic
versus an AI option. Now we’re at– since AI is
needed, which type is best? So AI tends to be better at– you know, humans are
good at certain things and AI’s are good
at certain things. When we think about the spectrum
of AI and how it can help out, there’s two kind of
useful categories. One is automation and the
other is augmentation. Automation tends to be– automation is just automating
a task essentially away. Augmentation is a way to
extend people’s abilities, but they retain agency. So for automation,
it’s very good when people don’t know
how to do something, people literally
can’t do something. So for example, if
you said process these millions of
rows of flight data, a human just can’t
really do that. Or if a task is boring,
repetitive, or dangerous something like maybe
measuring air quality in potentially toxic areas. Augmentation is often
preferred and better when people enjoy or feel
responsible for a task, like writing a birthday card. There are high stakes
situations or complicated personal preferences. In the flights example,
there are both high stakes– it’s both a high
stakes situation and they’re complicated
personal preferences. High stakes because
flights are expensive and getting to a family
wedding is important. Kristie flies out
tonight to a wedding. And she can attest that
it’s very important that she gets there on time
and not late in any way. So we got to make
sure we end on time. In terms of preferences,
there are many complicated personal preferences as everyone
likes to fly in their own way. One of the big divisive
ones for the flight’s team was overnight
flights, some people really, really love
overnight flights. Some people will
never, ever, ever go on an overnight
flight no matter what. And so they really wanted to
augment that person’s ability to make that decision
versus make it for them. So what did the
Google Flights team do to identify if AI adds
unique value to their product. They identified people’s
needs and behavior patterns. They defined AI success
in a people-centered way. They determined if the
scope of the problem needed AI and
determined that it did. And they decided on
the right type of AI to augment people’s ability
to make those decisions. And so once they
identified that need, it was time to translate
that need into data. And Kristie is going to tell
us about how they did that. KRISTIE FISHER: Thank you. So a big thing that we
learned synthesizing all of our knowledge across
Google about building people-centered AI is that it’s
often very tempting to look at the data that you have and
then say, huh, what kind of AI could I create from these data? But that is really
doing yourself and your future
users a disservice. The right thing to ask
is given my user need, what scope of data
would I need in order to then train an AI that could
actually meet that need well? So if you don’t take
a step back and go through that human
centered design process, you’re going to end
up with an AI that maybe works great
but doesn’t actually do what people need it to do. So to avoid that,
we want to make sure we are translating our
users’ needs into a data set that can be used
to train our AI. So how to think
about doing this? The first thing to
do is to ask yourself whether user’s needs align
clearly with AI outputs that an AI could
actually provide. Then you need to
think about mapping those outputs onto a data set
that you actually have access to. Then you need to
think about given the data set that I
need, can I source it in a way that is responsible
to make sure that I have no fairness or bias
issues in my data and to make sure that my
data set is robust enough to actually train a
model that will work. And then of course, you have
to do the really hard stuff that there’s a dozen other talks
on here today at I/O, which is build and train your model. But then as you do that
iteratively and tune your model over time to
get better and better, we’re going to challenge you
to do that in a way that really keeps people in mind and is
not just focused on getting to a certain statistic. So let’s start by thinking
about how to align users, people’s needs with data. So this is really simple. It’s really just taking a few
hours, talking with your team, thinking through it
yourself, going back to that research
you did with people, and thinking carefully about
all the different parts of the need of the person
who will be using this. Then thinking about,
given what they need, what could an AI output be
that could meet that need? So for example, on the flight’s
team, as Jess mentioned, the real big question
that people had was when do I need to book this? Is it going to go up tomorrow? Or can I wait and maybe get
paid first before I jump and buy that ticket? So in order to meet
the user’s needs, the AI output had to provide
first, an understanding of what the presenter price is. Is this current price
low, medium, or high? Then they also had to
say what do we think will happen in the future? Is it going to go up? Is it going to go down, or
is it going to stay the same? So in order to actually
meet the user’s need and help them buy a
flight at the right time, the AI output had to provide
these two key things. So once you understand
what the AI really needs to do to meet
the person’s need, then you can start thinking
about what sort of elements need to be in your data set in
order to then train an AI that can deliver those outputs. So as many of you probably
know, if you are experimenting with machine learning
in your own development, structured data sets
are a really common way to train models through
supervised learning. And structured data
sets contain examples. Examples are essentially
the rows of your data set. And those examples
contain features, which are the categories of
information in your data set. And then lastly,
you have labels. Labels are often
qualitative bits of data, and they’re often
generated by people. So as you think about
what data do you need, you need to be
thinking about, can I get a data set with
all of the features that would be required
to get an AI output that would meet my user’s need? And can I get all the labels
that I need to meet that need? So for example, for the
price team, or sorry, for the flight’s
team, if they only had price data about what
the historical price was but they didn’t know
what day it was today and what the price
was today, then that wouldn’t really
give them what they needed to meet people’s needs. People need to know both what
is the price today and what was the price in the past
in order for the AI output to give them what
they needed in terms of putting the price in
context and knowing what would happen in the future. So you really need to
think about do I have– can I get all of the
right features in my data set that could then lead
me to an AI output that would be valuable to people? And then once you
know all of the data that you need to
train your AI model, now you actually
have to find it. And this is no easy task. So for the flight’s
team, they already had a heuristic
based tool in place before they decided to try to
use AI for flight insights. And even then with
all this data that had been coming in for
many months and years, it still takes a lot of time
and effort as many of you know to actually
clean up that data and restructure it and make sure
that it’s in the proper format to then train a model. But for you, you
might not even have that huge pool of data sitting
around like Google does. So it’s going to take you
even more time and effort. If you’ve never worked with
machine learning and AI before, you might not realize just
how much time and effort this takes. But please know that
this is the best place to spend that time and
effort, because really making sure that you have the right
data that can then deliver the right AI outputs that can
then meet the person’s need is going to make
all the difference. So some key things
to ask yourself as you’re going on this
quest for your data is, is there a publicly available
data set that I could use? Or do I have to build
one from scratch? And then does my data set
contain enough example diversity to give me the
types of outputs that I need? So for example, again,
thinking of flights, if they’re going to
be able to provide meaningful insights for anyone
across the world looking for a flight, they need
millions of examples of historical flights,
including all the features we talked about where the
flight was originating from, where it was going,
how much it cost, when the ticket was
purchased, all of that stuff. And then lastly,
is your data set free of issues with privacy,
fairness, bias, et cetera? So each of those big
questions could of course, be a series of I/O talks. And there’s many specific talks
about these today and tomorrow as well and probably yesterday. So we’re not going to go
into all of the detail. But we just want to remind you
that we have a lot of resources on Google AI as well as provided
by the PAIR team and some tips in the guidebook for
building your own data set, making sure you’re using all the
possible resources that Google has made available to
make sure that you’re getting the best possible
data to train your model. And there’s one
tool you might not be aware of yet called
the Facets tool. The Facets tool allows
you to visualize what is inside of your
data set before you put it into your model
to train your ML. So what’s really helpful about
this is when you suddenly see all of your data
categorized and colored, it can be really obvious
if something is missing or if there’s a disproportionate
amount of one category of data versus another. The flight’s team was actually
able to catch a few errors itself by using this tool. So you finally
figured out what data you need and you spent all the
time getting the right data set and building and
training your model. Now you have to make sure
that you can continue to tweak and tune it to make it as
accurate and as reliable as it can impossibly be. And as you do this,
we’re going to give you a framework for
how we’d like you to think about us to do this in
a way that’s people centered. So that framework is this. If some metric for
our AI changes, the delta drops above or goes
below a certain threshold, then we’re going to
take a certain action. So what this meant for
the flights insights team was if bookings from this new
AI powered flight insights tool dropped below a
certain percentage, so people weren’t
using it, they weren’t finding it valuable
during the beta phase, then the flights team needed
to do something about it. They needed to then perhaps
adjust the confidence to make sure that only
the highest confident, highest confidence flight
insights predictions were actually making it in
to the flights insights UI. So what the flights team did to
translate user needs into data was first make sure that
the needs of the people who would be using
flight insights could be aligned with AI outputs. As we said, those AI outputs
are our current understanding of the existing price
today and a prediction for how that flight will
change in the future. Then they had to map those
AI outputs onto a training data set that they
could actually acquire, make sure that it
was sourced well and didn’t have any issues. And lastly, they had to
continue to tune and iterate on the model in a way
that was people centered. So now you might think we’ve
really done the hard work. We have built an AI that works. Like that is so hard. It can take months. It can take a team of
engineers, a really long time. Even with all the
tools that we’ve made available at Google
to democratize AI, it can be really hard. But unfortunately, there’s
still more you have to do. I’m sorry to tell you. Because if you build a great AI
but people don’t understand it, then you risk them
misunderstanding your model, using it incorrectly, or
simply ignoring your AI product altogether. This is something that
I think about every day because on the
Google Ads team we build really complex
AI driven forecasting tools that can help advertisers
make really hard decisions. But if they don’t understand
what went into this forecast, and how they’re meant to
use it and how confident they should be in it, then
they’re not really sure how to incorporate it
into their workflows. So we’re always having to work
really hard to make it better. So for you, what to think about
when you’re thinking about how to explain your AI system
to the people who are using it is first think
about what explanations are needed to build something
that we call calibrated trust. Calibrated trust means that
somebody does not under trust or over trust your AI. Under trusting is this
recommendation is crap, and I’m never going
to use that ever. Over trusting is ah,
well, if the AI says is, it must be the truth. I will take it no matter what. So with calibrated
trust, people know when they should use what the
AI gives them without question and when they should maybe
double check it or apply their own judgment. The next question
to consider when considering how to explain
your AI to the people using it is what level of technical
expertise do people have? Because people with higher
technical expertise may need or expect more information
about the AI in order to build that calibrated
trust compared with people who are non-experts. So for example, somebody
who is a neurologist who is using an AI system to help
them diagnose the patient, they’re going to really
want to make sure they understand all the
ins and outs of the AI and how it was
programmed and what data it was trained
on before they’re willing to trust it because
their patient’s life is in their hands. Meanwhile, if I’m
just using maps to find my way out of here so
I can get back to the office, I don’t need to know all
of the details of how Google Maps helped me get to
this direction and this route. And lastly, you need to
think about if and how to display your AI’s
prediction confidence. So most AI systems
can generate some sort of measure of how confident
they are in a given output or a given prediction. And how to visualize
that to people and whether they need
that information at all is something to think
really carefully about and something that we
give a lot of guidance on in the guidebook. So first, thinking about how
to build calibrated trust. Usually people do not
need to know everything. And sometimes full
disclosure is less helpful than just the
key pieces of information that are most important. So through doing
research and through doing lots of testing
and iterating, you will come to know what
information the people using your product need in order
to trust it and use it effectively. So typically, you’re
not going to give a full explanation
of here’s everything about how my AI system works. You’re going to give
a partial explanation. And typically, those
partial explanations should include what sources
of data were used by the FBI. This is important
because users need to understand when they
might have information that the system doesn’t have. So going back to that
neurologist example, they need to know
whether this prediction from the AI that this
person has a tumor is just based on lots of
imaging data from lots of people or whether it’s also taking
the very unique patient history of this one person
in front of me into account. Without knowing that what
the source of data is, it’s hard for that neurologist
to then trust the AI. And generally speaking, these
partial explanations of AI should include inputs,
outputs, capabilities, and limitations of the system. So this is what the AI
is expecting from you. This is what you can expect
to get back in return. And this is what it
simply can and can’t do. And as I mentioned before and
I’ll mention several more times before I give the
clicker back to Jess is that displaying the
confidence of the AI system is not always useful to help
the person use it effectively. So for the flight’s team, when
we think about what information the users of Flight
Insights needed in order to trust the system,
they needed to know what airlines and options
were or were not included. As Jess mentioned,
if I am totally willing to take overnight
flights in order to save some money,
I want to make sure that the AI knows that and is
taking all of those options into account. And through research,
the flights team learned as well that only
very confident predictions were going to actually be
useful to the people using flight insights. So if a prediction
was only about 50% confident that the flight
price would or would not go up, people didn’t want to see that. Didn’t seem much better than
what they were currently doing. So next is technical expertise. So thinking again about
flights, obviously, we can assume low
technical expertise. Whether even if you are
only flying once a year, you could still use
Flights Insights. It’s meant for people
all over the world who are in lots of
different geographies and who do fly a
lot, don’t fly a lot, so you can make sure that
this is understandable and in plain language that
lots of different people can understand. So understanding the low
expertise of the user, the flight’s team
built a framework for communicating in
plain language what the AI behind Flight
Insights was predicting. So they started off
building simple sentences that explained the
current state of the world and what the AI
predicted for the future. So for example,
prices are currently either low, medium, or high. And then what the AI
predicted would happen next. So to walk through a
couple of examples, let’s say prices
are currently high, and the system does not
have a high confidence prediction for what will happen
next if it’ll go up or down. In this case, the
Flight Insights tool might just say prices are high. I don’t have a good prediction
for what happens next. But just so you know for
context, prices are high. If the Flight Insights model did
have a good prediction for what would happen next,
it would again explain this to the
people in plain language. So prices are currently medium,
but they will generally go up. Or prices are currently low and
so today is a good day to book. As we already mentioned,
displaying model confidence is not always useful. And lots of different
types of displays exist and everyone
has trade-offs. We unfortunately aren’t able
to give you the out-of-the-box solution of this way of
displaying confidence for your AI will work for
everyone all of the time. What we can tell you is
here are some examples, here are the trade-offs
offs, and for you, you need to test and iterate
a lot to get it right. So some common ways of
displaying AI prediction confidence are categorical
in best, numeric, and data Datavis. So in categorical, you’re
simply chunking your predictions into really confident, kind
of confident, and not so confident. In N-best, you’re giving a
result and then you’re saying, here are some backups
just in case we’re wrong. So for example, here’s
a picture of Tokyo, but it might also be a picture
of New York or Los Angeles, or somewhere else. And numeric is simply
giving the number that this system has given you. I am 88% confident that
this recommendation is true. And Datavis, you
might show graphically what the AI’s prediction
is and then put some confidence intervals or
some error bars around it. And again, thinking
about the trade-offs and again about the users
likely technical expertise, keep in mind that
they don’t need to understand the AI as well as
you, the person that built it. They just need to understand it
enough to know what to do next. So when you think of something
like the numerical indicator, if you tell me that
you’re 88% confident I’m going to love this t-shirt,
is that good or bad? Do I need to know that
level of granularity? Is 86% that much better
or worse than 88? So those are the kinds
of things to think about. So again, for displaying
the AI prediction confidence for the flight’s
team, they decided not to give any kind
of indicator of how confident the system was. They simply decided
not to display it if it was not confident. So instead, their
design challenge was, how do we display
only these very confident predictions in a way that’s
simple and usable and elegant? So here’s what the Google
Flights UI looks like. And if we zoom in here,
here’s the indicator that again explains
in plain language what the AI is predicting. So here’s the current
price and here’s what we think will happen next. So to recap, you need to
build calibrated trust with your users, you need to
keep their expertise in mind when you’re creating
these explanations, and you need to decide and think
very carefully about if and how to display model confidence. So that’s how the flight team
built this insights tool. Jess, how did their
people-centered approach to AI design work? JESS HOLBROOK: Yeah, so
the team took insights from the guidebook, take a
people-centered approach. And we can’t share
exact numbers. But things have been
going really well. They get very positive feedback. And one of the most
important things to know is when you build in
a people-centered way, people will tell
you that you did. People will tell you
if you succeeded. There’ll be very
little guesswork. So to zoom back out from
the Google Flights example back to the general ideas,
what we’ve covered today is a few insights from
three of the chapters– how to identify if AI adds
unique value for your product or your use case,
how to then take that and translate that user
need into data needs and requirements
for your AI model, and then once you
have a functioning model, how to explain that
well to users in a way that they understand and
that they can feel in control of what they’re doing. Like we said at the
beginning, there are six chapters in
the guidebook total. Each has many of these
kind of insights in them. And so there’s just a ton
more for you to check out. The– you can check it out here. We hope everyone goes,
takes a look at it. We hope you use it. We hope it’s useful to you. Really importantly, we really
hope you send us feedback. There’s a big feedback
button along the whole thing because we want to
make something that’s useful for everyone out there. We really appreciate
your time today. If you want to keep
the conversation going, we’ll be at the Material Design
and Accessibility Sandbox at about 250 later on. So thank you very much. KRISTIE FISHER: Thank you. [APPLAUSE] [MUSIC PLAYING]


5 thoughts on “Designing Human-Centered AI Products (Google I/O’19)

  1. I didn't ready the whole book, but I guess it might be missing GDPR related issues when collecting data

  2. For the google flight case,
    1.Will it be helpful to review the flight companies algorithm first before building your own AI model?
    2. To build " calibrated trust", maybe we need to remove the Aura that AI is kind of prophetic of human's decision making behaviours. Only if one day, when people share the same " thoughts" with AI, AI could predict. People should still trust themselves to make decisions, INNT?

Leave a Reply

Your email address will not be published. Required fields are marked *