2019-06-01_New_Scientist

(singke) #1

W


HEN I was growing
up, nobody promised
me a flying car. But
I was promised an AI apocalypse.
Those shiny machines were
going to crush our skulls
underfoot, and we were all
going to welcome our new robot
overlords. Remember that?
Many people still seem to think
it is likely to happen. Well, just like
flying cars, it isn’t. But we might
still get a deadly AI nightmare.
It is just going to be a lot weirder
and more insidious than we
imagined in the innocent days of
Terminators and the Cybermen.
AI is simply software programs
that learn from data. That is why
engineers generally prefer to call
it machine learning. You give your
machine-learning program a giant
data set – say, every video on
YouTube – and it “learns” to find
patterns. Famously, when Google
unleashed an AI on YouTube
in 2012, it figured out how to
recognise cat faces. That sounds
pretty amazing until you discover
that Google’s AI also became
intrigued by spatulas oriented
at a 30 degree angle.
If you want to understand the
true dangers of AI, you have to
ponder those spatulas. It isn’t so
much that these programs are
working incorrectly, but that they
notice patterns we don’t. And
sometimes those patterns are
a lot more problematic than
an inadvertent spatula fetish.
When Microsoft released a
chatty AI bot named Tay on
Twitter, for example, it began
spouting Nazi slogans within
24 hours. Designed to learn
from conversations around
it on the social platform, it
quickly became racist.
A similar issue showed up in
crime-prediction algorithms that
police in Florida used. Software
flagged black people as more likely

to commit crimes than white
people, despite evidence to the
contrary. And when political
scientist Virginia Eubanks
investigated medical insurance
algorithms in the US, she
discovered an inherent bias
that made it harder for poor
people to get health coverage.
None of this should be
surprising to anyone who has
met a human and discovered
our propensity for prejudice.
AIs aren’t autonomous creatures
with agendas of their own. They
are learning from our data. Think
of AI as prostheses – extensions of
humanity, with slightly different
strengths and weaknesses.

Sometimes they suss out
patterns of bias in our data much
better than we do. Then, like the
obedient programs they are,
they act on those biases.
Google and Microsoft both
acknowledged this problem for
the first time in recent annual
reports to the Securities Exchange
Commission. AI, said Google, may
present “ethical, technological,
legal, and other challenges”.
Microsoft put it more simply:
“AI algorithms may be flawed.”
Taking human data out of the
equation doesn’t help – problems

often get worse when AIs learn
from other AIs. Algorithmic
trading between software bots
led to the trillion-dollar “flash
crash” in 2010, when the US stock
market plunged 998.5 points,
then recovered, within about
30 minutes.
YouTube engineers reported
recently that they fear what they
call the “inversion”. It is a scenario
in which the network becomes
so clogged with accounts run by
bots that AIs learn to dismiss real
visitors as fake.
The good news is that all these
failures mean we won’t have to
fight a robot army. The only way
we can prevent the AI apocalypse,
such as it is, will be to debug
ourselves. Already, people are
working on ways to correct for
racial and class bias in algorithms.
They are also thinking more
about the ethics of deploying
automation for sensitive tasks.
Google employees mounted
a protest when they discovered
that their employer was designing
a machine-learning algorithm for
surveillance drones that would
automatically identify enemy
targets. Under pressure, the
company stopped working
on the project. But plenty of
other companies are building
automation – complete with
unconscious human bias – into
their weapons. That is why the
European Union called on the
international community to
regulate what it calls “lethal
autonomous weapons systems”
currently in development.
There is a hard road ahead.
It is easier to obsess over
imaginary killer robots than
it is to undo decades of biased
data that we barely understand.
Still, this overwhelming task
offers a glimmer of hope. In the
end, improving our AI may also
improve humanity. ❚

24 | New Scientist | 1 June 2019


Views Columnist


PO

RT
RA
IT:^
JAN

E^ A

ND

ER
S;^ R

IGH

T:^ M

OV

IES

TO

RE
CO

LLE

CT
ION

LT
D/A

LA
MY

ST

OC

K^ P

HO

TO

This column will appear
monthly. Up next week:
botanist James Wong

Annalee Newitz is a science
journalist and author. Her
novel Autonomous won
the Lambda Literary Award
and she is the co-host of
the Hugo-nominated
podcast Our Opinions Are
Correct. You can follow her
@annaleen and her website
is techsploitation.com

Forget the AI apocalypse as you know it The true danger
of artificial intelligence is in its obsession with spatulas and
the biases it learns from us, writes Annalee Newitz

This changes everything


What are you reading?
P. Djèlí Clark’s The Black
God’s Drums, an alternate
history of the 19th century
Caribbean with mad
scientists and airships.

What are you
watching?
I just saw the gorgeous,
epic film The Wandering
Earth, where we turn our
planet into a spaceship.

What are you
working on?
I’m doing a lot of research
on ancient Roman toilets.

Annalee’s week


“ The AI nightmare
is going to be a lot
weirder than we
imagined in the
innocent days
of Terminators”
Free download pdf