0:00:00.739,0:00:04.861
So, I started my first job[br]as a computer programmer

0:00:04.885,0:00:06.841
in my very first year of college --

0:00:06.865,0:00:08.372
basically, as a teenager.

0:00:08.889,0:00:10.621
Soon after I started working,

0:00:10.645,0:00:12.255
writing software in a company,

0:00:12.799,0:00:16.434
a manager who worked at the company[br]came down to where I was,

0:00:16.458,0:00:17.726
and he whispered to me,

0:00:18.229,0:00:21.090
"Can he tell if I'm lying?"

0:00:21.806,0:00:23.883
There was nobody else in the room.

0:00:25.032,0:00:29.421
"Can who tell if you're lying?[br]And why are we whispering?"

0:00:30.266,0:00:33.373
The manager pointed[br]at the computer in the room.

0:00:33.397,0:00:36.493
"Can he tell if I'm lying?"

0:00:37.613,0:00:41.975
Well, that manager was having[br]an affair with the receptionist.

0:00:41.999,0:00:43.111
(Laughter)

0:00:43.135,0:00:44.901
And I was still a teenager.

0:00:45.447,0:00:47.466
So I whisper-shouted back to him,

0:00:47.490,0:00:51.114
"Yes, the computer can tell[br]if you're lying."

0:00:51.138,0:00:52.944
(Laughter)

0:00:52.968,0:00:55.891
Well, I laughed, but actually,[br]the laugh's on me.

0:00:55.915,0:00:59.183
Nowadays, there are computational systems

0:00:59.207,0:01:02.755
that can suss out[br]emotional states and even lying

0:01:02.779,0:01:04.823
from processing human faces.

0:01:05.248,0:01:09.401
Advertisers and even governments[br]are very interested.

0:01:10.319,0:01:12.181
I had become a computer programmer

0:01:12.205,0:01:15.318
because I was one of those kids[br]crazy about math and science.

0:01:15.942,0:01:19.050
But somewhere along the line[br]I'd learned about nuclear weapons,

0:01:19.074,0:01:22.026
and I'd gotten really concerned[br]with the ethics of science.

0:01:22.050,0:01:23.254
I was troubled.

0:01:23.278,0:01:25.919
However, because of family circumstances,

0:01:25.943,0:01:29.241
I also needed to start working[br]as soon as possible.

0:01:29.265,0:01:32.564
So I thought to myself, hey,[br]let me pick a technical field

0:01:32.588,0:01:34.384
where I can get a job easily

0:01:34.408,0:01:38.426
and where I don't have to deal[br]with any troublesome questions of ethics.

0:01:39.022,0:01:40.551
So I picked computers.

0:01:40.575,0:01:41.679
(Laughter)

0:01:41.703,0:01:45.113
Well, ha, ha, ha![br]All the laughs are on me.

0:01:45.137,0:01:47.891
Nowadays, computer scientists[br]are building platforms

0:01:47.915,0:01:52.124
that control what a billion[br]people see every day.

0:01:53.052,0:01:56.874
They're developing cars[br]that could decide who to run over.

0:01:57.707,0:02:00.920
They're even building machines, weapons,

0:02:00.944,0:02:03.229
that might kill human beings in war.

0:02:03.253,0:02:06.024
It's ethics all the way down.

0:02:07.183,0:02:09.241
Machine intelligence is here.

0:02:09.823,0:02:13.297
We're now using computation[br]to make all sort of decisions,

0:02:13.321,0:02:15.207
but also new kinds of decisions.

0:02:15.231,0:02:20.403
We're asking questions to computation[br]that have no single right answers,

0:02:20.427,0:02:21.629
that are subjective

0:02:21.653,0:02:23.978
and open-ended and value-laden.

0:02:24.002,0:02:25.760
We're asking questions like,

0:02:25.784,0:02:27.434
"Who should the company hire?"

0:02:28.096,0:02:30.855
"Which update from which friend[br]should you be shown?"

0:02:30.879,0:02:33.145
"Which convict is more[br]likely to reoffend?"

0:02:33.514,0:02:36.568
"Which news item or movie[br]should be recommended to people?"

0:02:36.592,0:02:39.964
Look, yes, we've been using[br]computers for a while,

0:02:39.988,0:02:41.505
but this is different.

0:02:41.529,0:02:43.596
This is a historical twist,

0:02:43.620,0:02:48.957
because we cannot anchor computation[br]for such subjective decisions

0:02:48.981,0:02:54.401
the way we can anchor computation[br]for flying airplanes, building bridges,

0:02:54.425,0:02:55.684
going to the moon.

0:02:56.449,0:02:59.708
Are airplanes safer?[br]Did the bridge sway and fall?

0:02:59.732,0:03:04.230
There, we have agreed-upon,[br]fairly clear benchmarks,

0:03:04.254,0:03:06.493
and we have laws of nature to guide us.

0:03:06.517,0:03:09.911
We have no such anchors and benchmarks

0:03:09.935,0:03:13.898
for decisions in messy human affairs.

0:03:13.922,0:03:18.159
To make things more complicated,[br]our software is getting more powerful,

0:03:18.183,0:03:21.956
but it's also getting less[br]transparent and more complex.

0:03:22.542,0:03:24.582
Recently, in the past decade,

0:03:24.606,0:03:27.335
complex algorithms[br]have made great strides.

0:03:27.359,0:03:29.349
They can recognize human faces.

0:03:29.985,0:03:32.040
They can decipher handwriting.

0:03:32.436,0:03:34.502
They can detect credit card fraud

0:03:34.526,0:03:35.715
and block spam

0:03:35.739,0:03:37.776
and they can translate between languages.

0:03:37.800,0:03:40.374
They can detect tumors in medical imaging.

0:03:40.398,0:03:42.603
They can beat humans in chess and Go.

0:03:43.264,0:03:47.768
Much of this progress comes[br]from a method called "machine learning."

0:03:48.175,0:03:51.362
Machine learning is different[br]than traditional programming,

0:03:51.386,0:03:54.971
where you give the computer[br]detailed, exact, painstaking instructions.

0:03:55.378,0:03:59.560
It's more like you take the system[br]and you feed it lots of data,

0:03:59.584,0:04:01.240
including unstructured data,

0:04:01.264,0:04:03.542
like the kind we generate[br]in our digital lives.

0:04:03.566,0:04:06.296
And the system learns[br]by churning through this data.

0:04:06.669,0:04:08.195
And also, crucially,

0:04:08.219,0:04:12.599
these systems don't operate[br]under a single-answer logic.

0:04:12.623,0:04:15.582
They don't produce a simple answer;[br]it's more probabilistic:

0:04:15.606,0:04:19.089
"This one is probably more like[br]what you're looking for."

0:04:20.023,0:04:23.093
Now, the upside is:[br]this method is really powerful.

0:04:23.117,0:04:25.193
The head of Google's AI systems called it,

0:04:25.217,0:04:27.414
"the unreasonable effectiveness of data."

0:04:27.791,0:04:29.144
The downside is,

0:04:29.738,0:04:32.809
we don't really understand[br]what the system learned.

0:04:32.833,0:04:34.420
In fact, that's its power.

0:04:34.946,0:04:38.744
This is less like giving[br]instructions to a computer;

0:04:39.200,0:04:43.264
it's more like training[br]a puppy-machine-creature

0:04:43.288,0:04:45.659
we don't really understand or control.

0:04:46.362,0:04:47.913
So this is our problem.

0:04:48.427,0:04:52.689
It's a problem when this artificial[br]intelligence system gets things wrong.

0:04:52.713,0:04:56.253
It's also a problem[br]when it gets things right,

0:04:56.277,0:04:59.905
because we don't even know which is which[br]when it's a subjective problem.

0:04:59.929,0:05:02.268
We don't know what this thing is thinking.

0:05:03.493,0:05:07.176
So, consider a hiring algorithm --

0:05:08.123,0:05:12.434
a system used to hire people,[br]using machine-learning systems.

0:05:13.052,0:05:16.631
Such a system would have been trained[br]on previous employees' data

0:05:16.655,0:05:19.246
and instructed to find and hire

0:05:19.270,0:05:22.308
people like the existing[br]high performers in the company.

0:05:22.814,0:05:23.967
Sounds good.

0:05:23.991,0:05:25.990
I once attended a conference

0:05:26.014,0:05:29.139
that brought together[br]human resources managers and executives,

0:05:29.163,0:05:30.369
high-level people,

0:05:30.393,0:05:31.952
using such systems in hiring.

0:05:31.976,0:05:33.622
They were super excited.

0:05:33.646,0:05:38.299
They thought that this would make hiring[br]more objective, less biased,

0:05:38.323,0:05:41.323
and give women[br]and minorities a better shot

0:05:41.347,0:05:43.535
against biased human managers.

0:05:43.559,0:05:46.402
And look -- human hiring is biased.

0:05:47.099,0:05:48.284
I know.

0:05:48.308,0:05:51.313
I mean, in one of my early jobs[br]as a programmer,

0:05:51.337,0:05:55.205
my immediate manager would sometimes[br]come down to where I was

0:05:55.229,0:05:58.982
really early in the morning[br]or really late in the afternoon,

0:05:59.006,0:06:02.068
and she'd say, "Zeynep,[br]let's go to lunch!"

0:06:02.724,0:06:04.891
I'd be puzzled by the weird timing.

0:06:04.915,0:06:07.044
It's 4pm. Lunch?

0:06:07.068,0:06:10.162
I was broke, so free lunch. I always went.

0:06:10.618,0:06:12.685
I later realized what was happening.

0:06:12.709,0:06:17.255
My immediate managers[br]had not confessed to their higher-ups

0:06:17.279,0:06:20.392
that the programmer they hired[br]for a serious job was a teen girl

0:06:20.416,0:06:24.346
who wore jeans and sneakers to work.

0:06:25.174,0:06:27.376
I was doing a good job,[br]I just looked wrong

0:06:27.400,0:06:29.099
and was the wrong age and gender.

0:06:29.123,0:06:32.469
So hiring in a gender- and race-blind way

0:06:32.493,0:06:34.358
certainly sounds good to me.

0:06:35.031,0:06:38.372
But with these systems,[br]it is more complicated, and here's why:

0:06:38.968,0:06:44.759
Currently, computational systems[br]can infer all sorts of things about you

0:06:44.783,0:06:46.655
from your digital crumbs,

0:06:46.679,0:06:49.012
even if you have not[br]disclosed those things.

0:06:49.506,0:06:52.433
They can infer your sexual orientation,

0:06:52.994,0:06:54.300
your personality traits,

0:06:54.859,0:06:56.232
your political leanings.

0:06:56.830,0:07:00.515
They have predictive power[br]with high levels of accuracy.

0:07:01.362,0:07:03.940
Remember -- for things[br]you haven't even disclosed.

0:07:03.964,0:07:05.555
This is inference.

0:07:05.579,0:07:08.840
I have a friend who developed[br]such computational systems

0:07:08.864,0:07:12.505
to predict the likelihood[br]of clinical or postpartum depression

0:07:12.529,0:07:13.945
from social media data.

0:07:14.676,0:07:16.103
The results are impressive.

0:07:16.492,0:07:19.849
Her system can predict[br]the likelihood of depression

0:07:19.873,0:07:23.776
months before the onset of any symptoms --

0:07:23.800,0:07:25.173
months before.

0:07:25.197,0:07:27.443
No symptoms, there's prediction.

0:07:27.467,0:07:32.279
She hopes it will be used[br]for early intervention. Great!

0:07:32.911,0:07:34.951
But now put this in the context of hiring.

0:07:36.027,0:07:39.073
So at this human resources[br]managers conference,

0:07:39.097,0:07:43.806
I approached a high-level manager[br]in a very large company,

0:07:43.830,0:07:48.408
and I said to her, "Look,[br]what if, unbeknownst to you,

0:07:48.432,0:07:54.981
your system is weeding out people[br]with high future likelihood of depression?

0:07:55.761,0:07:59.137
They're not depressed now,[br]just maybe in the future, more likely.

0:07:59.923,0:08:03.329
What if it's weeding out women[br]more likely to be pregnant

0:08:03.353,0:08:05.939
in the next year or two[br]but aren't pregnant now?

0:08:06.844,0:08:12.480
What if it's hiring aggressive people[br]because that's your workplace culture?"

0:08:13.173,0:08:15.864
You can't tell this by looking[br]at gender breakdowns.

0:08:15.888,0:08:17.390
Those may be balanced.

0:08:17.414,0:08:20.971
And since this is machine learning,[br]not traditional coding,

0:08:20.995,0:08:25.902
there is no variable there[br]labeled "higher risk of depression,"

0:08:25.926,0:08:27.759
"higher risk of pregnancy,"

0:08:27.783,0:08:29.517
"aggressive guy scale."

0:08:29.995,0:08:33.674
Not only do you not know[br]what your system is selecting on,

0:08:33.698,0:08:36.021
you don't even know[br]where to begin to look.

0:08:36.045,0:08:37.291
It's a black box.

0:08:37.315,0:08:40.122
It has predictive power,[br]but you don't understand it.

0:08:40.486,0:08:42.855
"What safeguards," I asked, "do you have

0:08:42.879,0:08:46.552
to make sure that your black box[br]isn't doing something shady?"

0:08:48.863,0:08:52.741
She looked at me as if I had[br]just stepped on 10 puppy tails.

0:08:52.765,0:08:54.013
(Laughter)

0:08:54.037,0:08:56.078
She stared at me and she said,

0:08:56.556,0:09:00.889
"I don't want to hear[br]another word about this."

0:09:01.458,0:09:03.492
And she turned around and walked away.

0:09:04.064,0:09:05.550
Mind you -- she wasn't rude.

0:09:05.574,0:09:11.882
It was clearly: what I don't know[br]isn't my problem, go away, death stare.

0:09:11.906,0:09:13.152
(Laughter)

0:09:13.862,0:09:17.701
Look, such a system[br]may even be less biased

0:09:17.725,0:09:19.828
than human managers in some ways.

0:09:19.852,0:09:21.998
And it could make monetary sense.

0:09:22.573,0:09:24.223
But it could also lead

0:09:24.247,0:09:28.995
to a steady but stealthy[br]shutting out of the job market

0:09:29.019,0:09:31.312
of people with higher risk of depression.

0:09:31.753,0:09:34.349
Is this the kind of society[br]we want to build,

0:09:34.373,0:09:36.658
without even knowing we've done this,

0:09:36.682,0:09:40.646
because we turned decision-making[br]to machines we don't totally understand?

0:09:41.265,0:09:42.723
Another problem is this:

0:09:43.314,0:09:47.766
these systems are often trained[br]on data generated by our actions,

0:09:47.790,0:09:49.606
human imprints.

0:09:50.188,0:09:53.996
Well, they could just be[br]reflecting our biases,

0:09:54.020,0:09:57.613
and these systems[br]could be picking up on our biases

0:09:57.637,0:09:58.950
and amplifying them

0:09:58.974,0:10:00.392
and showing them back to us,

0:10:00.416,0:10:01.878
while we're telling ourselves,

0:10:01.902,0:10:05.019
"We're just doing objective,[br]neutral computation."

0:10:06.314,0:10:08.991
Researchers found that on Google,

0:10:10.134,0:10:15.447
women are less likely than men[br]to be shown job ads for high-paying jobs.

0:10:16.463,0:10:18.993
And searching for African-American names

0:10:19.017,0:10:23.723
is more likely to bring up ads[br]suggesting criminal history,

0:10:23.747,0:10:25.314
even when there is none.

0:10:26.693,0:10:30.242
Such hidden biases[br]and black-box algorithms

0:10:30.266,0:10:34.239
that researchers uncover sometimes[br]but sometimes we don't know,

0:10:34.263,0:10:36.924
can have life-altering consequences.

0:10:37.958,0:10:42.117
In Wisconsin, a defendant[br]was sentenced to six years in prison

0:10:42.141,0:10:43.496
for evading the police.

0:10:44.824,0:10:46.010
You may not know this,

0:10:46.034,0:10:50.032
but algorithms are increasingly used[br]in parole and sentencing decisions.

0:10:50.056,0:10:53.011
He wanted to know:[br]How is this score calculated?

0:10:53.795,0:10:55.460
It's a commercial black box.

0:10:55.484,0:10:59.689
The company refused to have its algorithm[br]be challenged in open court.

0:11:00.396,0:11:05.928
But ProPublica, an investigative[br]nonprofit, audited that very algorithm

0:11:05.952,0:11:07.968
with what public data they could find,

0:11:07.992,0:11:10.308
and found that its outcomes were biased

0:11:10.332,0:11:13.961
and its predictive power[br]was dismal, barely better than chance,

0:11:13.985,0:11:18.401
and it was wrongly labeling[br]black defendants as future criminals

0:11:18.425,0:11:22.320
at twice the rate of white defendants.

0:11:23.891,0:11:25.455
So, consider this case:

0:11:26.103,0:11:29.955
This woman was late[br]picking up her godsister

0:11:29.979,0:11:32.054
from a school in Broward County, Florida,

0:11:32.757,0:11:35.113
running down the street[br]with a friend of hers.

0:11:35.137,0:11:39.236
They spotted an unlocked kid's bike[br]and a scooter on a porch

0:11:39.260,0:11:40.892
and foolishly jumped on it.

0:11:40.916,0:11:43.515
As they were speeding off,[br]a woman came out and said,

0:11:43.539,0:11:45.744
"Hey! That's my kid's bike!"

0:11:45.768,0:11:49.062
They dropped it, they walked away,[br]but they were arrested.

0:11:49.086,0:11:52.723
She was wrong, she was foolish,[br]but she was also just 18.

0:11:52.747,0:11:55.291
She had a couple of juvenile misdemeanors.

0:11:55.808,0:12:00.993
Meanwhile, that man had been arrested[br]for shoplifting in Home Depot --

0:12:01.017,0:12:03.941
85 dollars' worth of stuff,[br]a similar petty crime.

0:12:04.766,0:12:09.325
But he had two prior[br]armed robbery convictions.

0:12:09.955,0:12:13.437
But the algorithm scored her[br]as high risk, and not him.

0:12:14.746,0:12:18.620
Two years later, ProPublica found[br]that she had not reoffended.

0:12:18.644,0:12:21.194
It was just hard to get a job[br]for her with her record.

0:12:21.218,0:12:23.294
He, on the other hand, did reoffend

0:12:23.318,0:12:27.154
and is now serving an eight-year[br]prison term for a later crime.

0:12:28.088,0:12:31.457
Clearly, we need to audit our black boxes

0:12:31.481,0:12:34.096
and not have them have[br]this kind of unchecked power.

0:12:34.120,0:12:36.999
(Applause)

0:12:38.087,0:12:42.329
Audits are great and important,[br]but they don't solve all our problems.

0:12:42.353,0:12:45.101
Take Facebook's powerful[br]news feed algorithm --

0:12:45.125,0:12:49.968
you know, the one that ranks everything[br]and decides what to show you

0:12:49.992,0:12:52.276
from all the friends and pages you follow.

0:12:52.898,0:12:55.173
Should you be shown another baby picture?

0:12:55.197,0:12:56.393
(Laughter)

0:12:56.417,0:12:59.013
A sullen note from an acquaintance?

0:12:59.449,0:13:01.305
An important but difficult news item?

0:13:01.329,0:13:02.811
There's no right answer.

0:13:02.835,0:13:05.494
Facebook optimizes[br]for engagement on the site:

0:13:05.518,0:13:06.933
likes, shares, comments.

0:13:08.168,0:13:10.864
In August of 2014,

0:13:10.888,0:13:13.550
protests broke out in Ferguson, Missouri,

0:13:13.574,0:13:17.991
after the killing of an African-American[br]teenager by a white police officer,

0:13:18.015,0:13:19.585
under murky circumstances.

0:13:19.974,0:13:21.981
The news of the protests was all over

0:13:22.005,0:13:24.690
my algorithmically[br]unfiltered Twitter feed,

0:13:24.714,0:13:26.664
but nowhere on my Facebook.

0:13:27.182,0:13:28.916
Was it my Facebook friends?

0:13:28.940,0:13:30.972
I disabled Facebook's algorithm,

0:13:31.472,0:13:34.320
which is hard because Facebook[br]keeps wanting to make you

0:13:34.344,0:13:36.380
come under the algorithm's control,

0:13:36.404,0:13:38.642
and saw that my friends[br]were talking about it.

0:13:38.666,0:13:41.175
It's just that the algorithm[br]wasn't showing it to me.

0:13:41.199,0:13:44.241
I researched this and found[br]this was a widespread problem.

0:13:44.265,0:13:48.078
The story of Ferguson[br]wasn't algorithm-friendly.

0:13:48.102,0:13:49.273
It's not "likable."

0:13:49.297,0:13:50.849
Who's going to click on "like?"

0:13:51.500,0:13:53.706
It's not even easy to comment on.

0:13:53.730,0:13:55.101
Without likes and comments,

0:13:55.125,0:13:58.417
the algorithm was likely showing it[br]to even fewer people,

0:13:58.441,0:13:59.983
so we didn't get to see this.

0:14:00.946,0:14:02.174
Instead, that week,

0:14:02.198,0:14:04.496
Facebook's algorithm highlighted this,

0:14:04.520,0:14:06.746
which is the ALS Ice Bucket Challenge.

0:14:06.770,0:14:10.512
Worthy cause; dump ice water,[br]donate to charity, fine.

0:14:10.536,0:14:12.440
But it was super algorithm-friendly.

0:14:13.219,0:14:15.832
The machine made this decision for us.

0:14:15.856,0:14:19.353
A very important[br]but difficult conversation

0:14:19.377,0:14:20.932
might have been smothered,

0:14:20.956,0:14:23.652
had Facebook been the only channel.

0:14:24.117,0:14:27.914
Now, finally, these systems[br]can also be wrong

0:14:27.938,0:14:30.674
in ways that don't resemble human systems.

0:14:30.698,0:14:33.620
Do you guys remember Watson,[br]IBM's machine-intelligence system

0:14:33.644,0:14:36.772
that wiped the floor[br]with human contestants on Jeopardy?

0:14:37.131,0:14:38.559
It was a great player.

0:14:38.583,0:14:42.152
But then, for Final Jeopardy,[br]Watson was asked this question:

0:14:42.659,0:14:45.591
"Its largest airport is named[br]for a World War II hero,

0:14:45.615,0:14:47.867
its second-largest[br]for a World War II battle."

0:14:47.891,0:14:49.269
(Hums Final Jeopardy music)

0:14:49.582,0:14:50.764
Chicago.

0:14:50.788,0:14:52.158
The two humans got it right.

0:14:52.697,0:14:57.045
Watson, on the other hand,[br]answered "Toronto" --

0:14:57.069,0:14:58.887
for a US city category!

0:14:59.596,0:15:02.497
The impressive system also made an error

0:15:02.521,0:15:06.172
that a human would never make,[br]a second-grader wouldn't make.

0:15:06.823,0:15:09.932
Our machine intelligence can fail

0:15:09.956,0:15:13.056
in ways that don't fit[br]error patterns of humans,

0:15:13.080,0:15:16.030
in ways we won't expect[br]and be prepared for.

0:15:16.054,0:15:19.692
It'd be lousy not to get a job[br]one is qualified for,

0:15:19.716,0:15:23.443
but it would triple suck[br]if it was because of stack overflow

0:15:23.467,0:15:24.899
in some subroutine.

0:15:24.923,0:15:26.502
(Laughter)

0:15:26.526,0:15:29.312
In May of 2010,

0:15:29.336,0:15:33.380
a flash crash on Wall Street[br]fueled by a feedback loop

0:15:33.404,0:15:36.432
in Wall Street's "sell" algorithm

0:15:36.456,0:15:40.640
wiped a trillion dollars[br]of value in 36 minutes.

0:15:41.722,0:15:43.909
I don't even want to think[br]what "error" means

0:15:43.933,0:15:47.522
in the context of lethal[br]autonomous weapons.

0:15:49.894,0:15:53.684
So yes, humans have always made biases.

0:15:53.708,0:15:55.884
Decision makers and gatekeepers,

0:15:55.908,0:15:59.401
in courts, in news, in war ...

0:15:59.425,0:16:02.463
they make mistakes;[br]but that's exactly my point.

0:16:02.487,0:16:06.008
We cannot escape[br]these difficult questions.

0:16:06.596,0:16:10.112
We cannot outsource[br]our responsibilities to machines.

0:16:10.676,0:16:14.884
(Applause)

0:16:17.089,0:16:21.536
Artificial intelligence does not give us[br]a "Get out of ethics free" card.

0:16:22.742,0:16:26.123
Data scientist Fred Benenson[br]calls this math-washing.

0:16:26.147,0:16:27.536
We need the opposite.

0:16:27.560,0:16:32.948
We need to cultivate algorithm suspicion,[br]scrutiny and investigation.

0:16:33.380,0:16:36.578
We need to make sure we have[br]algorithmic accountability,

0:16:36.602,0:16:39.047
auditing and meaningful transparency.

0:16:39.380,0:16:42.614
We need to accept[br]that bringing math and computation

0:16:42.638,0:16:45.608
to messy, value-laden human affairs

0:16:45.632,0:16:48.016
does not bring objectivity;

0:16:48.040,0:16:51.673
rather, the complexity of human affairs[br]invades the algorithms.

0:16:52.148,0:16:55.635
Yes, we can and we should use computation

0:16:55.659,0:16:57.673
to help us make better decisions.

0:16:57.697,0:17:03.029
But we have to own up[br]to our moral responsibility to judgment,

0:17:03.053,0:17:05.871
and use algorithms within that framework,

0:17:05.895,0:17:10.830
not as a means to abdicate[br]and outsource our responsibilities

0:17:10.854,0:17:13.308
to one another as human to human.

0:17:13.807,0:17:16.416
Machine intelligence is here.

0:17:16.440,0:17:19.861
That means we must hold on ever tighter

0:17:19.885,0:17:22.032
to human values and human ethics.

0:17:22.056,0:17:23.210
Thank you.

0:17:23.234,0:17:28.254
(Applause)