How to pronounce "fret"

So when people voice fears of artificial intelligence,

very often, they invoke images of humanoid robots run amok.

You know? Terminator?

You know, that might be something to consider,

but that's a distant threat.

Or, we fret about digital surveillance

with metaphors from the past.

"1984," George Orwell's "1984,"

it's hitting the bestseller lists again.

It's a great book,

but it's not the correct dystopia for the 21st century.

What we need to fear most

is not what artificial intelligence will do to us on its own,

but how the people in power will use artificial intelligence

to control us and to manipulate us

in novel, sometimes hidden,

subtle and unexpected ways.

Much of the technology

that threatens our freedom and our dignity in the near-term future

is being developed by companies

in the business of capturing and selling our data and our attention

to advertisers and others:

Facebook, Google, Amazon,

Alibaba, Tencent.

Now, artificial intelligence has started bolstering their business as well.

And it may seem like artificial intelligence

is just the next thing after online ads.

It's not.

It's a jump in category.

It's a whole different world,

and it has great potential.

It could accelerate our understanding of many areas of study and research.

But to paraphrase a famous Hollywood philosopher,

"With prodigious potential comes prodigious risk."

Now let's look at a basic fact of our digital lives, online ads.

Right? We kind of dismiss them.

They seem crude, ineffective.

We've all had the experience of being followed on the web

by an ad based on something we searched or read.

You know, you look up a pair of boots

and for a week, those boots are following you around everywhere you go.

Even after you succumb and buy them, they're still following you around.

We're kind of inured to that kind of basic, cheap manipulation.

We roll our eyes and we think, "You know what? These things don't work."

Except, online,

the digital technologies are not just ads.

Now, to understand that, let's think of a physical world example.

You know how, at the checkout counters at supermarkets, near the cashier,

there's candy and gum at the eye level of kids?

That's designed to make them whine at their parents

just as the parents are about to sort of check out.

Now, that's a persuasion architecture.

It's not nice, but it kind of works.

That's why you see it in every supermarket.

Now, in the physical world,

such persuasion architectures are kind of limited,

because you can only put so many things by the cashier. Right?

And the candy and gum, it's the same for everyone,

even though it mostly works

only for people who have whiny little humans beside them.

In the physical world, we live with those limitations.

In the digital world, though,

persuasion architectures can be built at the scale of billions

and they can target, infer, understand

and be deployed at individuals

one by one

by figuring out your weaknesses,

and they can be sent to everyone's phone private screen,

so it's not visible to us.

And that's different.

And that's just one of the basic things that artificial intelligence can do.

Now, let's take an example.

Let's say you want to sell plane tickets to Vegas. Right?

So in the old world, you could think of some demographics to target

based on experience and what you can guess.

You might try to advertise to, oh,

men between the ages of 25 and 35,

or people who have a high limit on their credit card,

or retired couples. Right?

That's what you would do in the past.

With big data and machine learning,

that's not how it works anymore.

So to imagine that,

think of all the data that Facebook has on you:

every status update you ever typed,

every Messenger conversation,

every place you logged in from,

all your photographs that you uploaded there.

If you start typing something and change your mind and delete it,

Facebook keeps those and analyzes them, too.

Increasingly, it tries to match you with your offline data.

It also purchases a lot of data from data brokers.

It could be everything from your financial records

to a good chunk of your browsing history.

Right? In the US, such data is routinely collected,

collated and sold.

In Europe, they have tougher rules.

So what happens then is,

by churning through all that data, these machine-learning algorithms --

that's why they're called learning algorithms --

they learn to understand the characteristics of people

who purchased tickets to Vegas before.

When they learn this from existing data,

they also learn how to apply this to new people.

So if they're presented with a new person,

they can classify whether that person is likely to buy a ticket to Vegas or not.

Fine. You're thinking, an offer to buy tickets to Vegas.

I can ignore that.

But the problem isn't that.

The problem is,

we no longer really understand how these complex algorithms work.

We don't understand how they're doing this categorization.

It's giant matrices, thousands of rows and columns,

maybe millions of rows and columns,

and not the programmers

and not anybody who looks at it,

even if you have all the data,

understands anymore how exactly it's operating

any more than you'd know what I was thinking right now

if you were shown a cross section of my brain.

It's like we're not programming anymore,

we're growing intelligence that we don't truly understand.

And these things only work if there's an enormous amount of data,

so they also encourage deep surveillance on all of us

so that the machine learning algorithms can work.

That's why Facebook wants to collect all the data it can about you.

The algorithms work better.

So let's push that Vegas example a bit.

What if the system that we do not understand

was picking up that it's easier to sell Vegas tickets

to people who are bipolar and about to enter the manic phase.

Such people tend to become overspenders, compulsive gamblers.

They could do this, and you'd have no clue that's what they were picking up on.

I gave this example to a bunch of computer scientists once

and afterwards, one of them came up to me.

He was troubled and he said, "That's why I couldn't publish it."

I was like, "Couldn't publish what?"

He had tried to see whether you can indeed figure out the onset of mania

from social media posts before clinical symptoms,

and it had worked,

and it had worked very well,

and he had no idea how it worked or what it was picking up on.

Now, the problem isn't solved if he doesn't publish it,

because there are already companies

that are developing this kind of technology,

and a lot of the stuff is just off the shelf.

This is not very difficult anymore.

Do you ever go on YouTube meaning to watch one video

and an hour later you've watched 27?

You know how YouTube has this column on the right

that says, "Up next"

and it autoplays something?

It's an algorithm

picking what it thinks that you might be interested in

and maybe not find on your own.

It's not a human editor.

It's what algorithms do.

It picks up on what you have watched and what people like you have watched,

and infers that that must be what you're interested in,

what you want more of,

and just shows you more.

It sounds like a benign and useful feature,

except when it isn't.

So in 2016, I attended rallies of then-candidate Donald Trump

to study as a scholar the movement supporting him.

I study social movements, so I was studying it, too.

And then I wanted to write something about one of his rallies,

so I watched it a few times on YouTube.

YouTube started recommending to me

and autoplaying to me white supremacist videos

in increasing order of extremism.

If I watched one,

it served up one even more extreme

and autoplayed that one, too.

If you watch Hillary Clinton or Bernie Sanders content,

YouTube recommends and autoplays conspiracy left,

and it goes downhill from there.

Well, you might be thinking, this is politics, but it's not.

This isn't about politics.

This is just the algorithm figuring out human behavior.

I once watched a video about vegetarianism on YouTube

and YouTube recommended and autoplayed a video about being vegan.

It's like you're never hardcore enough for YouTube.

(Laughter)

So what's going on?

Now, YouTube's algorithm is proprietary,

but here's what I think is going on.

The algorithm has figured out

that if you can entice people

into thinking that you can show them something more hardcore,

they're more likely to stay on the site

watching video after video going down that rabbit hole

while Google serves them ads.

Now, with nobody minding the ethics of the store,

these sites can profile people

who are Jew haters,

who think that Jews are parasites

and who have such explicit anti-Semitic content,

and let you target them with ads.

They can also mobilize algorithms

to find for you look-alike audiences,

people who do not have such explicit anti-Semitic content on their profile

but who the algorithm detects may be susceptible to such messages,

and lets you target them with ads, too.

Now, this may sound like an implausible example,

but this is real.

ProPublica investigated this

and found that you can indeed do this on Facebook,

and Facebook helpfully offered up suggestions

on how to broaden that audience.

BuzzFeed tried it for Google, and very quickly they found,

yep, you can do it on Google, too.

And it wasn't even expensive.

The ProPublica reporter spent about 30 dollars

to target this category.

So last year, Donald Trump's social media manager disclosed

that they were using Facebook dark posts to demobilize people,

not to persuade them,

but to convince them not to vote at all.

And to do that, they targeted specifically,

for example, African-American men in key cities like Philadelphia,

and I'm going to read exactly what he said.

I'm quoting.

They were using "nonpublic posts

whose viewership the campaign controls

so that only the people we want to see it see it.

We modeled this.

It will dramatically affect her ability to turn these people out."

What's in those dark posts?

We have no idea.

Facebook won't tell us.

So Facebook also algorithmically arranges the posts

that your friends put on Facebook, or the pages you follow.

It doesn't show you everything chronologically.

It puts the order in the way that the algorithm thinks will entice you

to stay on the site longer.

Now, so this has a lot of consequences.

You may be thinking somebody is snubbing you on Facebook.

The algorithm may never be showing your post to them.

The algorithm is prioritizing some of them and burying the others.

Experiments show

that what the algorithm picks to show you can affect your emotions.

But that's not all.

It also affects political behavior.

So in 2010, in the midterm elections,

Facebook did an experiment on 61 million people in the US

that was disclosed after the fact.

So some people were shown, "Today is election day,"

the simpler one,

and some people were shown the one with that tiny tweak

with those little thumbnails

of your friends who clicked on "I voted."

This simple tweak.

OK? So the pictures were the only change,

and that post shown just once

turned out an additional 340,000 voters

in that election,

according to this research

as confirmed by the voter rolls.

A fluke? No.

Because in 2012, they repeated the same experiment.

And that time,

that civic message shown just once

turned out an additional 270,000 voters.

For reference, the 2016 US presidential election

was decided by about 100,000 votes.

Now, Facebook can also very easily infer what your politics are,

even if you've never disclosed them on the site.

Right? These algorithms can do that quite easily.

What if a platform with that kind of power

decides to turn out supporters of one candidate over the other?

How would we even know about it?

Now, we started from someplace seemingly innocuous --

online adds following us around --

and we've landed someplace else.

As a public and as citizens,

we no longer know if we're seeing the same information

or what anybody else is seeing,

and without a common basis of information,

little by little,

public debate is becoming impossible,

and we're just at the beginning stages of this.

These algorithms can quite easily infer

things like your people's ethnicity,

religious and political views, personality traits,

intelligence, happiness, use of addictive substances,

parental separation, age and genders,

just from Facebook likes.

These algorithms can identify protesters

even if their faces are partially concealed.

These algorithms may be able to detect people's sexual orientation

just from their dating profile pictures.

Now, these are probabilistic guesses,

so they're not going to be 100 percent right,

but I don't see the powerful resisting the temptation to use these technologies

just because there are some false positives,

which will of course create a whole other layer of problems.

Imagine what a state can do

with the immense amount of data it has on its citizens.

China is already using face detection technology

to identify and arrest people.

And here's the tragedy:

we're building this infrastructure of surveillance authoritarianism

merely to get people to click on ads.

And this won't be Orwell's authoritarianism.

This isn't "1984."

Now, if authoritarianism is using overt fear to terrorize us,

we'll all be scared, but we'll know it,

we'll hate it and we'll resist it.

But if the people in power are using these algorithms

to quietly watch us,

to judge us and to nudge us,

to predict and identify the troublemakers and the rebels,

to deploy persuasion architectures at scale

and to manipulate individuals one by one

using their personal, individual weaknesses and vulnerabilities,

and if they're doing it at scale

through our private screens

so that we don't even know

what our fellow citizens and neighbors are seeing,

that authoritarianism will envelop us like a spider's web

and we may not even know we're in it.

So Facebook's market capitalization

is approaching half a trillion dollars.

It's because it works great as a persuasion architecture.

But the structure of that architecture

is the same whether you're selling shoes

or whether you're selling politics.

The algorithms do not know the difference.

The same algorithms set loose upon us

to make us more pliable for ads

are also organizing our political, personal and social information flows,

and that's what's got to change.

Now, don't get me wrong,

we use digital platforms because they provide us with great value.

I use Facebook to keep in touch with friends and family around the world.

I've written about how crucial social media is for social movements.

I have studied how these technologies can be used

to circumvent censorship around the world.

But it's not that the people who run, you know, Facebook or Google

are maliciously and deliberately trying

to make the country or the world more polarized

and encourage extremism.

I read the many well-intentioned statements

that these people put out.

But it's not the intent or the statements people in technology make that matter,

it's the structures and business models they're building.

And that's the core of the problem.

Either Facebook is a giant con of half a trillion dollars

and ads don't work on the site,

it doesn't work as a persuasion architecture,

or its power of influence is of great concern.

It's either one or the other.

It's similar for Google, too.

So what can we do?

This needs to change.

Now, I can't offer a simple recipe,

because we need to restructure

the whole way our digital technology operates.

Everything from the way technology is developed

to the way the incentives, economic and otherwise,

are built into the system.

We have to face and try to deal with

the lack of transparency created by the proprietary algorithms,

the structural challenge of machine learning's opacity,

all this indiscriminate data that's being collected about us.

We have a big task in front of us.

We have to mobilize our technology,

our creativity

and yes, our politics

so that we can build artificial intelligence

that supports us in our human goals

but that is also constrained by our human values.

And I understand this won't be easy.

We might not even easily agree on what those terms mean.

But if we take seriously

how these systems that we depend on for so much operate,

I don't see how we can postpone this conversation anymore.

These structures

are organizing how we function

and they're controlling

what we can and we cannot do.

And many of these ad-financed platforms,

they boast that they're free.

In this context, it means that we are the product that's being sold.

We need a digital economy

where our data and our attention

is not for sale to the highest-bidding authoritarian or demagogue.

(Applause)

So to go back to that Hollywood paraphrase,

we do want the prodigious potential

of artificial intelligence and digital technology to blossom,

but for that, we must face this prodigious menace,

open-eyed and now.

Thank you.

(Applause)