CES 2021: Experts weigh in on the importance of diversity in building AI

In programming new artificial
intelligence software, diversity
among developers translates
into better business outcomes,
according to AI experts
speaking at this year’s CES,
the Consumer Technology
Association’s annual technology
and media tradeshow.
That is because without diverse
views in development, the
technology only learns how to
think from a more limited point
of view. This can have serious
— and sometimes expensive —
repercussions as unexpected
problems are discovered after
the software rolls out, the experts
noted.
“The most important thing we
all need to remember is that
inclusive inputs lead to inclusive
outputs,” said Annie JeanBaptiste, head of production
inclusion at Google LLC. “It’s
really important to have
perspectives that have been
historically underrepresented
or at the margins of the creation
of AI and make sure we have

inputs that actually reflect the
diversity of our world today.”
Kimberly Sterling, a senior
director at medical device
and digital health company
ResMed, pointed to lessons
learned during the pandemic
about challenges for diverse
populations in healthcare as
one example of why more
diversity is needed in a number
of fields. This is especially true
for software developers who
are trying to make accurate AI
models, she said.
“We are on the precipice of
this amazing technology in AI,
but if we don’t have diverse
data sets based on diverse
populations really reflective on
the heterogeneity in the world,
we end up with bad prediction
models,” Sterling said.
Taniya Mishra, founder and CEO
of AI company SureStart, noted
that when she first worked on
voice recognition technology
in the early 2000s, the most
commonly available data sets
were samples of “newsreader

speech,” comprising Caucasians
speaking in standard American
accents. Due to a lack of foreign
accents available, the AI had
issues trying to decipher anyone
not talking in a very specific way.
“Although data sets in voice
recognition systems have made
significant strides in the last
decade, the issues of not being
understood to persist,” Mishra
said, citing voices of children
and the elderly as two common
examples.
When asked by the moderator
to highlight examples of ways
to integrate inclusivity within
AI, Jean-Baptiste spoke about
the effort that Google put into
making sure its AI-powered
Google Assistant did not respond
to user queries with harmful or
biased language.
“We knew there were potential
groups that weren’t reflected
on the product design team
that needed to have a seat
at the table — thinking about
race, gender, ability, age, socioeconomic status and sexual

orientation,” Jean-Baptiste said.
To remedy this, Google
brought in employees from
underrepresented backgrounds
to run the AI through adversarial
testing, essentially trying to
break the product before
it launches, Jean-Baptiste
explained. Doing this, the team
not only uncovered negative
things they did not want the AI to
say, but it also proactively added
positive cultural references.
As a result, the final product
received very few inquiries that
the company had to act on, JeanBaptiste said.
“This was a testament to building
for everyone and with everyone,”
she said. “It’s important to have
as many voices as possible
because inclusion really fuels
innovation, and if we want
to see the best of the best in
terms of technology we need to
make sure that voices that have
been at the margins are really
included.”