22 Scientific American, June 2019VENTURES
THE BUSINESS OF INNOVATIONWade Roush is the host and producer of Soonish, a podcast
about technology, culture, curiosity and the future. He
is a co-founder of the podcast collective Hub & Spoke and
a freelance reporter for print, online and radio outlets,
such as MIT Technology Review, Xconomy, WBUR and WHYY.Illustration by Jay BendtSafe Words for
Our AI Friends
Virtual assistants are getting smarter.
Let’s think about how that will play out
By Wade RoushIn their latest iteration, Apple’s popular AirPods wireless ear
buds let you activate Siri, Apple’s AI assistant, simply by say
ing, “Hey, Siri,” just as you can with your iPhone. With the orig
inal AirPods, a physical tap on one AirPod would bring up Siri,
but the voice command is simpler. And it takes us one step
closer to a world where we can talk to our AIs and they to us
anywhere, anytime.
It’s a technology we’ve been anticipating for decades. From
the Enterprise computer on the original Star Trek (1966–1969)
to HAL 9000 in 2001: A Space Odyssey (1968) to Samantha in
Spike Jonze’s Her (2013), science fiction has shown us all manner
of disembodied AI helpmates who can answer our questions,
carry out our orders or even provide emotional intimacy.
With the emergence of AIs like Siri, Google Assistant, Ama
zon’s Alexa and Microsoft’s Cortana, the idea is now a lot less fic
tional. I’d genuinely miss Alexa if I couldn’t ask her to supply
weather forecasts, keep my shopping list, control the lights in
my house, and play podcasts and radio.
But AI assistants aren’t yet omnipresent, and they aren’t all
that smart. Their arrival in our ear canals, plus some stunning
recent progress in AI research, will change all that. In Silicon Val
ley, Google and OpenAI, a nonprofit research company, have been
racing to apply advances in an area called unsupervised learning.
Their latest language models cull existing texts on the Web to gen
erate coherent, humanlike responses in questionanswering and
textcompletion tasks. Within a couple of years these models will
make AI assistants dramatically more capable and talkative.
And that means it’s time to ask whether we really want AIs
whispering in our ears all day—and if so, what conditions and
controls we’d like to see implemented alongside them.
In last month’s Ventures column, I looked at the ways Face
book’s seemingly benign plan to connect people with one anoth
er went off the rails, resulting in a system of mass surveillance
and manipulation. The same thing could happen with AI assis
tants if we don’t insist on basic protections in advance. Let me
suggest a few:Privacy. Inevitably the smarts of our AIs will reside in the cloud,
on servers owned by tech giants such as Amazon, Apple, Google
and Microsoft. So our interactions with AIs should be encrypt
ed end to end—unreadable even by the companies—and the
records should be automatically deleted after a short period.
Transparency. AI providers must be up front about how they
are handling our data, how customer behavior feeds back into
improvements in the system, and how they are making money,
without burying the details in unreadable, 50page enduser
license agreements.
Security and reliability. We will engage with our AI assistants in
our homes, vehicles and workplaces across numerous WiFi and
(soon) 5G networks. We will be relying on them for advice, sugges
tions and answers, at the same time we will be giving them real
world tasks such as monitoring the performance of our applianc
es and the safety of our homes. We will need high avail ability, and
every link in the communications chain must be hackerproof.
Trustworthiness. The same unsupervised learning algorithms
that generate coherent conversation could be coopted to generate
fake or misleading content—which is part of the reason OpenAI is
not yet releasing its powerful new language models to the outside
world. When we ask our AIs for answers, we’ll need assurances
that they are drawing on accurate data from trusted sources.
Autonomy. AI assistants should exist to give us more agency
over our lives, not less. It would be a disaster for everyone if they
morphed into vehicles for selling us things, stealing our atten
tion or stoking our anxieties.If the giant AI providers are allowed to selfregulate in these
areas, the result will surely be more Facebookstyle fiascoes. The
push for protections will have to come from us, the users, and
our representatives in government. After all, no one wants “Hey,
Siri,” to turn into “Bye, Siri.”JOIN THE CONVERSATION ONLINE
Visit Scientific American on Facebook and Twitter
or send a letter to the editor: [email protected]