Are We Automating Racism?

Vox
31 Mar 202122:54

Summary

TLDRDas Skript beleuchtet die Herausforderungen von künstlicher Intelligenz (KI) und Algorithmenbezug, insbesondere in Bezug auf rassistische Tendenzen. Es zeigt, wie KI-Systeme, die von menschlichen Entscheidungen und Vorurteilen beeinflusst sind, zu unerwarteten und unfairen Ergebnissen führen können. Das Video diskutiert Beispiele wie das Twitter-Bildzuschneiden, das rassistisch wirken könnte, und stellt die Notwendigkeit von Überwachung und ethischer Ausrichtung von KI-Technologien her. Es betont auch die Bedeutung der Fragestellung und der Verantwortung der Schöpfer von KI-Werkzeugen gegenüber der Gesellschaft.

Takeaways

  • 🤖 Künstliche Intelligenz (KI) kann diskriminierend sein, selbst wenn sie neutrale Absichten hat und keine menschliche Vorurteile enthält.
  • 📈 Datengetriebene Systeme werden zunehmend in unserem Leben eine größere Rolle spielen, aber ihre Fehlfunktionen betreffen nicht alle gleich.
  • 🖼️ Algorithmen können aufgrund von Datenmangel an Vielfalt oder fehlerhaften Beispielen rassistisch wirken, selbst wenn das nicht beabsichtigt ist.
  • 👀 KI-Systeme lernen von Beispielen, die von Menschen ausgewählt, gekennzeichnet und abgeleitet wurden.
  • 🔍 Das Problem von KI-Bias ist groß und schwierig, aber es bedeutet nicht, dass wir auf maschinelles Lernen verzichten müssen.
  • 📊 Eine der Vorteile von Computer-Bias im Vergleich zu menschlichem Bias ist, dass man ihn leicht messen und verfolgen kann.
  • 🌐 Die Auswahl und Verwendung von Algorithmen hängt von Ressourcen und Macht ab und wer die Interessen durch ein prädiktives Modell vertreten werden soll.
  • 📋 Modelle sollten dokumentiert werden, einschließlich ethischer Bedenken, wie der beabsichtigte Gebrauch, die Herkunft der Daten und Anweisungen zur Bewertung der Systemleistung.
  • 🚫 Es ist wichtig, vorsichtig zu sein, wo und wie wir KI-Systeme einsetzen und welche menschliche Überwachung wir über diese Systeme haben.
  • 🧐 Die Auswahl der Daten und das Labeling beeinflussen die KI-Modelle erheblich und können zu unerwünschten Ergebnissen führen, wenn nicht sorgfältig gemacht.
  • 🤔 Die Frage, welche Technologien entwickelt und wie sie in unserer Welt eingesetzt werden, reduziert sich auf Machtdynamiken und die Interessen derjenigen, die die Fragen stellen.
  • 📝 Die Verantwortung der Designer und Programmierer geht über die Rechenschaft gegenüber ihren Kunden hinaus und betrifft auch die Auswirkungen, die diese Werkzeuge in der Welt haben.

Q & A

  • Was ist das Hauptthema des Transcripts?

    -Das Hauptthema des Transcripts ist die Diskussion über den algorithmischen Bias und die Auswirkungen von künstlicher Intelligenz (KI) auf die Gesellschaft, insbesondere in Bezug auf Diskriminierung und Vorurteile.

  • Wie kann ein Bildschirm- oder Fotocropping-Algorithmus diskriminierend wirken?

    -Ein Bildschirm- oder Fotocropping-Algorithmus kann diskriminierend wirken, indem er bei der Auswahl des zu zeigenden Bildausschnitts zwischen verschiedenen Gesichtern eine Ungleichheit aufweist, z.B. indem er voreingenommen weiße Gesichter bevorzugt.

  • Was ist ein 'Saliency Prediction Model'?

    -Ein 'Saliency Prediction Model' ist eine Art KI-Software, die darauf abzielt, die wichtigen Elemente in einem Bild zu identifizieren. Es basiert auf den Daten, die durch das Beobachten von menschlichen Augenbewegungen bei der Betrachtung von Bildern gesammelt wurden.

  • Wie kann die Auswahl von Daten, die zur Schulung eines Modells verwendet werden, zu Bias führen?

    -Die Auswahl von Daten kann zu Bias führen, wenn sie nicht repräsentativ für die verschiedenen Bevölkerungsgruppen ist. Wenn beispielsweise ein Datensatz viele Katzenbilder enthält, aber nur wenige dunkelhäutige Gesichter, kann das Modell lernen, Katzen als wichtiger zu erkennen als bestimmte Menschen.

  • Was ist ein Beispiel für die Verwendung von KI in der Gesundheitsversorgung, die zu unerwünschten Ergebnissen führen kann?

    -Im Text wird ein Beispiel genannt, bei dem ein Programm zur Identifizierung von Patienten mit chronischen Krankheiten zur Zusatzversorgung verwendet wurde. Das Programm basierte auf den Kosten der Patienten, was jedoch dazu führte, dass Patienten mit dunklerer Hautfarbe aufgrund von Kostenunterscheiden, die mit institutionalisierter Rassismus und mangelnder Versorgung zu tun hatten, nicht angemessen identifiziert wurden.

  • Wie können Modelle zur KI-Entwicklung verbessert werden, um Bias zu reduzieren?

    -Modelle zur KI-Entwicklung können verbessert werden, indem sie mit einer Vielzahl von Daten trainiert werden, die eine breitere Repräsentanz der Bevölkerung widerspiegeln. Auch die Transparenz über die Datenherkunft und die Bezeichnung von Daten ist wichtig. Ein weiteres Werkzeug zur Verbesserung ist die sogenannte 'Modellkartei', die wichtige Informationen über das Modell und seine potenziellen ethischen Bedenken dokumentiert.

  • Was ist die Rolle von menschlicher Subjektivität in der KI-Entwicklung?

    -Menschliche Subjektivität spielt eine große Rolle in der KI-Entwicklung, da Entscheidungen über die Auswahl und Beschriftung von Daten, die zur Schulung des Modells verwendet werden, von Menschen getroffen werden. Diese Entscheidungen können unbewusst Bias in den Algorithmen einfügen.

  • Wie kann die Verwendung von KI-Technologien in der Gesellschaft ethische und gesellschaftliche Herausforderungen hervorrufen?

    -Die Verwendung von KI-Technologien kann ethische und gesellschaftliche Herausforderungen hervorrufen, da sie potenziell Diskriminierung und Ungerechtigkeit verfestigen können. Es ist wichtig, über die Auswirkungen von KI-Entscheidungen nachzudenken und sicherzustellen, dass sie nicht auf der Grundlage von Rasse, Geschlecht oder sozialem Status diskriminieren.

  • Was ist der Vorteil der Identifizierung von Bias in KI-Systemen gegenüber menschlichem Bias?

    -Der Vorteil der Identifizierung von Bias in KI-Systemen ist, dass sie messbar und nachverfolgbar ist und es ermöglicht, an dem Modell zu arbeiten, um faire Ergebnisse zu erzielen, wenn man dazu motiviert ist.

  • Welche Rolle spielen Ressourcen und Macht bei der Entscheidung darüber, welche Technologien entwickelt und wie sie in unserer Welt eingesetzt werden?

    -Ressourcen und Macht sind entscheidend, da sie bestimmen, wessen Interessen von einem prädiktiven Modell bedient werden und welche Fragen gestellt werden. Die Machtstrukturen legen fest, auf welchem Weg die Technologie eingesetzt wird und wer davon profitiert oder benachteiligt wird.

  • Was ist die Verantwortung von Designern und Programmierern in Bezug auf die Auswirkungen der von ihnen entwickelten Werkzeuge in der Welt?

    -Designer und Programmierer müssen sich fragen, ob sie nur ihren Clients gegenüber verantwortlich sind oder auch gegenüber der gesamten politischen Gemeinschaft. Sie müssen sich bewusst sein, welche Auswirkungen ihre Tools haben können und ob sie bereit sind, für diese Auswirkungen zu haften.

  • Was ist der Unterschied zwischen dem Bias in einem Computer und dem Bias beim Menschen?

    -Der Unterschied liegt darin, dass der Bias in einem Computer leichter gemessen und verfolgt werden kann, während der menschliche Bias oft komplexer zu identifizieren und zu quantifizieren ist.

  • Wie kann die Transparenz in Bezug auf die Entwicklung von KI-Modellen verbessert werden?

    -Die Transparenz kann verbessert werden, indem Unternehmen interne Dokumentationskulturen wie 'Modellkarteien' einführen, die wichtige Informationen über die Funktionsweise des Modells, die Datenherkunft und die ethischen Bedenken liefern.

Outlines

00:00

🤖 Künstliche Intelligenz und rassistische Entscheidungen

Der erste Absatz spricht über die Diskrepanz zwischen menschlicher Wahrnehmung und der eines Computers, der Menschen nur als Daten sieht. Es wird diskutiert, dass Daten-getriebene Systeme immer wichtiger werden und normalerweise gut funktionieren, aber wenn sie fehlschlagen, dies nicht auf alle gleichermaßen trifft. Ein Beispiel ist das rassistische Verhalten eines Algorithmus, der bei der Wahl von Gesichtern zwischen verschiedenen Personen diskriminiert. Die Diskussion umfasst auch die Bedeutung von rassistischen Auswirkungen und wie man mit ihnen umgeht, selbst wenn die Absichten neutral sind.

05:01

📈 Systematische Tests auf künstliche Intelligenz-Bias

In diesem Absatz wird ein öffentlicher Test von algorithmischem Bias durchgeführt, indem extreme vertikale Bilder von zwei Personen hochgeladen werden, um zu sehen, welcher von ihnen von einem Bildzuschneide-Algorithmus bevorzugt wird. Die Diskussion umfasst die Bedeutung von Testbarkeit, die Tatsache, dass es möglich ist, den Algorithmus sofort zu testen, und die Beobachtung, dass der Algorithmus möglicherweise weiße Gesichter bevorzugt. Es wird auch auf die Notwendigkeit einer systematischen Testung mit Hunderten von Fotos eingegangen, um zu einer Schlussfolgerung zu kommen.

10:02

🔍 Saliency Prediction Models und ihre Auswirkungen

Der dritte Absatz konzentriert sich auf die Funktionsweise von Saliency Prediction Models, die verwendet werden, um zu entscheiden, was in einem Bild wichtig ist. Es wird erläutert, wie diese Modelle mithilfe von Datensätzen, die menschliche Blickbewegungen dokumentieren, trainiert werden. Die Diskussion umfasst auch die Herausforderungen, die mit der Rückverfolgung von Entscheidungen verbunden sind, die von solchen Modellen getroffen werden, und wie die Daten, die zur Schulung verwendet werden, die Leistung des Modells beeinflussen. Beispiele für die Verwendung von Daten wie Kriminalitätsdaten und medizinischen Kennzahlen werden gegeben, um zu zeigen, wie Bias in den Daten die Ergebnisse beeinflussen kann.

15:04

💰 Die Auswirkungen von Algorithmen auf das Gesundheitswesen

Dieser Absatz behandelt ein Beispiel für die Anwendung von Algorithmen im Gesundheitswesen und wie diese fehlschlagen können, wenn sie auf Kosten statt auf der tatsächlichen Gesundheit basieren. Es wird erklärt, wie ein Algorithmus Patienten anhand ihrer Gesundheitskosten bewertet, anstatt anhand ihrer Krankheitsbilder, was zu rassistischen Ergebnissen führen kann. Die Diskussion umfasst auch die Bedeutung der Wahl der Variablen, die zur Definition von Hochrisikopatienten verwendet werden, und wie diese Entscheidungen die Ergebnisse beeinflussen können.

20:05

🛡️ Ethik und Verantwortung in der künstlichen Intelligenz

Der letzte Absatz spricht über die Herausforderungen, die mit der Verantwortung und der Ethik in Bezug auf künstliche Intelligenz verbunden sind. Es wird diskutiert, dass die Branche wenig reguliert ist und wie interne Dokumentationskulturen wie 'Model Cards' dazu beitragen können, Transparenz und Verantwortlichkeit zu fördern. Die Diskussion umfasst auch die Notwendigkeit, die Verwendung von Algorithmen kritisch zu hinterfragen und zu entscheiden, ob sie überhaupt eingesetzt werden sollten. Es wird betont, dass es um die Macht und Ressourcen geht, die bestimmen, welche Technologien entwickelt und wie sie in unserer Welt eingesetzt werden.

Mindmap

Keywords

💡Algorithmic Bias

Algorithmic Bias bezieht sich auf die systematische Verzerrung oder Diskriminierung, die durch die Verwendung von Algorithmen entsteht. Im Video wird dies durch Tests an einem Bildzuschneiden-Algorithmus von Twitter illustriert, der tendenziell weiße Gesichter gegenüber dunkeren Gesichtern bevorzugt. Dies zeigt, wie Algorithmen, die auf menschlichen Entscheidungen und Daten basieren, diese Biases widerspiegeln können.

💡Data-Driven Systems

Data-Driven Systems sind Systeme, die Entscheidungen oder Vorhersagen auf der Analyse von Daten beziehen. Im Video wird betont, dass diese Systeme immer wichtiger im Leben der Menschen werden, aber ihre Fehlfunktionen nicht gleichmäßig auf alle wirken, was auf ein grundsätzliches Problem in der Gestaltung dieser Systeme hinweist.

💡Machine Learning

Machine Learning ist ein Teil der Künstlichen Intelligenz, bei dem Computerprogramme lernen, Muster in Daten zu erkennen, ohne dass sie explizit programmiert werden. Im Video wird gezeigt, wie Maschinenlernen durch das Lernen von Beispielen funktioniert, die von Menschen gelabelt, ausgewählt und abgeleitet werden, was zu einer Wiederspiegelung menschlicher Vorurteile führen kann.

💡Racial Bias

Racial Bias ist ein Vorurteil oder eine Diskriminierung aufgrund der Rasse. Im Video wird diskutiert, dass selbst wenn die Absichten neutral sind, die Ergebnisse einer Technologie diskriminierend sein können. Dies wird anhand von Beispielen gezeigt, wie etwa der Tatsache, dass ein Algorithmus weiße Gesichter bei einem Bildzuschnitt bevorzugt.

💡Saliency Prediction Model

Ein Saliency Prediction Model ist ein Typ von Algorithmus, der versucht, die wichtigsten Elemente in einem Bild zu identifizieren. In dem Video wird dies durch die Analyse von Fotos von Twitter-Nutzern demonstriert, bei denen der Algorithmus tendenziell weiße Gesichter als salienter erkennt, was auf eine potenzielle Racial Bias hinweist.

💡Representation in Data Sets

Die Vertretung in Datenmengen bezieht sich auf die Vielfalt und den Umfang, in dem verschiedene Gruppen in den zur Modellierung und Training von KI verwendeten Daten enthalten sind. Im Video wird kritisiert, dass Datenmengen oft nicht repräsentativ sind, was zu einer unzureichenden Leistung von KI-Modellen bei unterrepräsentierten Gruppen führen kann.

💡Bias in Healthcare Algorithms

Bias in Healthcare Algorithms zeigt sich, wenn medizinische Entscheidungen oder Behandlungspläne aufgrund von fehlerhaften oder unzureichenden Daten diskriminierend wirken. Im Video wird ein Beispiel gegeben, bei dem ein Algorithmus Patienten anhand ihrer Kostenrisiken bewertete, was zu einer ungleichen Betreuung von Patienten unterschiedlicher Rassen führte.

💡Model Cards

Model Cards sind eine Art Dokumentation, die von Google vorgeschlagen wurde, um Informationen über KI-Modelle bereitzustellen, einschließlich ihrer Funktionsweise, beabsichtigten Verwendung und ethischer Bedenken. Im Video wird dies als eine Methode zur Verbesserung der Transparenz und Verantwortlichkeit von KI-Systemen beschrieben.

💡Ethical Concerns in AI

Ethical Concerns in AI beziehen sich auf die moralischen Fragen und Bedenken, die sich aus der Entwicklung und Verwendung von künstlicher Intelligenz ergeben. Im Video wird auf Probleme wie Racial Bias und die Notwendigkeit der Überwachung von KI-Systemen eingegangen, um sicherzustellen, dass sie fair und diskriminierungsfrei sind.

💡Human Oversight

Human Oversight bezieht sich auf die Überwachung und Kontrolle von KI-Systemen durch menschliche Experten, um sicherzustellen, dass sie ethisch und verantwortungsvoll handeln. Im Video wird betont, dass menschliche Überwachung notwendig ist, um Biases in AI zu identifizieren und zu korrigieren.

💡Power Dynamics in AI

Power Dynamics in AI beschreiben die Machtverteilung und -ausübung bei der Entscheidung, welche Technologien entwickelt und eingesetzt werden. Im Video wird darauf hingewiesen, dass es wichtig ist, die Interessen derjenigen zu berücksichtigen, die von AI-Systemen betroffen sind, anstatt nur die Interessen derjenigen, die diese Systeme erstellen und nutzen.

Highlights

Machines can exhibit bias even without human malice, raising questions about the fairness of data-driven systems.

A public test of algorithmic bias using images of Mitch McConnell and Barack Obama demonstrated potential racial bias in image cropping algorithms.

The idea of tech neutrality is debunked as biases can be embedded in AI systems, leading to discriminatory outcomes.

AI systems learn from human-labeled examples, which can perpetuate existing biases if not properly addressed.

The saliency prediction model used by Twitter for image cropping was tested and found to potentially favor white faces.

The composition of training data sets for AI is crucial; unrepresentative data can lead to biased algorithms.

An example of biased AI is in healthcare algorithms that use cost as a proxy for health risk, which can disadvantage certain demographic groups.

The healthcare algorithm case study shows how AI decisions can operate invisibly, impacting people without their knowledge.

Machine learning as an industry is largely unregulated, with no mandatory reporting of performance metrics or evaluation results.

Model Cards are a new documentation effort to provide transparency about how AI models work and their ethical considerations.

Bias in AI is measurable and can be tracked, offering a path toward achieving fair outcomes through model adjustment.

The deployment of AI technologies raises questions about which algorithms should be used and for what purposes.

The power dynamics in technology development determine whose interests are served by predictive models and which questions are asked.

Designers and programmers of AI tools must consider their accountability not just to clients but to the broader societal impact of their creations.

The challenge in AI ethics is not only in removing bias from algorithms but also in deciding which algorithms are necessary and beneficial.

Ruha Benjamin, a professor at Princeton University, discusses the systemic nature of bias in AI and its implications on society.

AI systems can reinforce societal biases if the data they learn from is not diverse or representative of various demographics.

Transcripts

play00:02

Maybe we-- if you guys could stand over--

play00:04

Is it okay if they stand over here?

play00:06

- Yeah. - Um, actually.

play00:08

Christophe, if you can get even lower.

play00:12

- Okay. - ( shutter clicks )

play00:13

This is Lee and this is Christophe.

play00:15

They're two of the hosts of this show.

play00:18

But to a machine, they're not people.

play00:21

This is just pixels. It's just data.

play00:23

A machine shouldn't have a reason to prefer

play00:25

one of these guys over the other.

play00:27

And yet, as you'll see in a second, it does.

play00:31

It feels weird to call a machine racist,

play00:36

but I really can't explain-- I can't explain what just happened.

play00:41

Data-driven systems are becoming a bigger and bigger part of our lives,

play00:45

and they work well a lot of the time.

play00:47

- But when they fail... - Once again, it's the white guy.

play00:51

When they fail, they're not failing on everyone equally.

play00:54

If I go back right now...

play00:58

Ruha Benjamin: You can have neutral intentions.

play01:00

You can have good intentions.

play01:03

And the outcomes can still be discriminatory.

play01:05

Whether you want to call that machine racist

play01:07

or you want to call the outcome racist,

play01:09

we have a problem.

play01:16

( theme music playing )

play01:23

I was scrolling through my Twitter feed a while back

play01:26

and I kept seeing tweets that look like this.

play01:29

Two of the same picture of Republican senator Mitch McConnell smiling,

play01:33

or sometimes it would be four pictures

play01:36

of the same random stock photo guy.

play01:39

And I didn't really know what was going on,

play01:42

but it turns out that this was a big public test of algorithmic bias.

play01:47

Because it turns out that these aren't pictures of just Mitch McConnell.

play01:50

They're pictures of Mitch McConnell and...

play01:54

- Barack Obama. - Lee: Oh, wow.

play01:57

So people were uploading

play01:58

these really extreme vertical images

play02:00

to basically force this image cropping algorithm

play02:03

to choose one of these faces.

play02:05

People were alleging that there's a racial bias here.

play02:08

But I think what's so interesting about this particular algorithm

play02:12

is that it is so testable for the public.

play02:15

It's something that we could test right now if we wanted to.

play02:19

- Let's do it. - You guys wanna do it?

play02:21

Okay. Here we go.

play02:26

So, Twitter does offer you options to crop your own image.

play02:30

But if you don't use those,

play02:32

it uses an automatic cropping algorithm.

play02:37

- Wow. There it is. - Whoa. Wow.

play02:39

That's crazy.

play02:41

Christophe, it likes you.

play02:43

Okay, let's try the other-- the happy one.

play02:44

Lee: Wow.

play02:48

- Unbelievable. Oh, wow. - Both times.

play02:53

So, do you guys think this machine is racist?

play02:58

The only other theory I possibly have

play03:00

is if the algorithm prioritizes white faces

play03:04

because it can pick them up quicker, for whatever reason,

play03:07

against whatever background.

play03:09

Immediately, it looks through the image

play03:11

and tries to scan for a face.

play03:13

Why is it always finding the white face first?

play03:16

Joss: With this picture, I think someone could argue

play03:19

that the lighting makes Christophe's face more sharp.

play03:24

I still would love to do

play03:26

a little bit more systematic testing on this.

play03:29

I think maybe hundreds of photos

play03:32

could allow us to draw a conclusion.

play03:34

I have downloaded a bunch of photos

play03:36

from a site called Generated Photos.

play03:39

These people do not exist. They were a creation of AI.

play03:43

And I went through, I pulled a bunch

play03:46

that I think will give us

play03:47

a pretty decent way to test this.

play03:50

So, Christophe, I wonder if you would be willing to help me out with that.

play03:54

You want me to tweet hundreds of photos?

play03:57

- ( Lee laughs ) - Joss: Exactly.

play03:59

I'm down. Sure, I've got time.

play04:04

Okay.

play04:05

( music playing )

play04:21

There may be some people who take issue with the idea

play04:24

that machines can be racist

play04:26

without a human brain or malicious intent.

play04:29

But such a narrow definition of racism

play04:32

really misses a lot of what's going on.

play04:34

I want to read a quote that responds to that idea.

play04:36

It says, "Robots are not sentient beings, sure,

play04:40

but racism flourishes well beyond hate-filled hearts.

play04:43

No malice needed, no "N" word required,

play04:46

just a lack of concern for how the past shapes the present."

play04:50

I'm going now to speak to the author of those words, Ruha Benjamin.

play04:54

She's a professor of African-American Studies at Princeton University.

play05:00

When did you first become concerned

play05:02

that automated systems, AI, could be biased?

play05:06

A few years ago, I noticed these headlines

play05:09

and hot takes about so-called racist and sexist robots.

play05:13

There was a viral video in which two friends were in a hotel bathroom

play05:18

and they were trying to use an automated soap dispenser.

play05:21

Black hand, nothing. Larry, go.

play05:28

Black hand, nothing.

play05:30

And although they seem funny

play05:32

and they kind of get us to chuckle,

play05:34

the question is, are similar design processes

play05:38

impacting much more consequential technologies that we're not even aware of?

play05:44

When the early news controversies came along maybe 10 years ago,

play05:49

people were surprised by the fact that they showed a racial bias.

play05:54

Why do you think people were surprised?

play05:55

Part of it is a deep attachment and commitment

play05:59

to this idea of tech neutrality.

play06:02

People-- I think because life is so complicated

play06:04

and our social world is so messy--

play06:07

really cling on to something that will save us,

play06:10

and a way of making decisions that's not drenched

play06:14

in the muck of all of human subjectivity,

play06:19

human prejudice and frailty.

play06:21

We want it so much to be true.

play06:22

We want it so much to be true, you know?

play06:24

And the danger is that we don't question it.

play06:27

And still we continue to have, you know, so-called glitches

play06:33

when it comes to race and skin complexion.

play06:36

And I don't think that they're glitches.

play06:38

It's a systemic issue in the truest sense of the word.

play06:41

It has to do with our computer systems and the process of design.

play06:47

Joss: AI can seem pretty abstract sometimes.

play06:50

So we built this to help explain

play06:52

how machine learning works and what can go wrong.

play06:55

This black box is the part of the system that we interact with.

play06:59

It's the software that decides which dating profiles we might like,

play07:02

how much a rideshare should cost,

play07:04

or how a photo should be cropped on Twitter.

play07:06

We just see a device making a decision.

play07:08

Or more accurately, a prediction.

play07:11

What we don't see is all of the human decisions

play07:13

that went into the design of that technology.

play07:17

Now, it's true that when you're dealing with AI,

play07:19

that means that the code in this box

play07:20

wasn't all written directly by humans,

play07:22

but by machine-learning algorithms

play07:25

that find complex patterns in data.

play07:27

But they don't just spontaneously learn things from the world.

play07:30

They're learning from examples.

play07:33

Examples that are labeled by people,

play07:35

selected by people,

play07:37

and derived from people, too.

play07:40

See, these machines and their predictions,

play07:42

they're not separate from us or from our biases

play07:44

or from our history,

play07:46

which we've seen in headline after headline

play07:48

for the past 10 years.

play07:51

We're using the face-tracking software,

play07:54

so it's supposed to follow me as I move.

play07:56

As you can see, I do this-- no following.

play08:01

Not really-- not really following me.

play08:03

- Wanda, if you would, please? - Sure.

play08:11

In 2010, the top hit

play08:14

when you did a search for "black girls,"

play08:15

80% of what you found

play08:17

on the first page of results was all porn sites.

play08:20

Google is apologizing after its photo software

play08:23

labeled two African-Americans gorillas.

play08:27

Microsoft is shutting down

play08:28

its new artificial intelligent bot

play08:31

after Twitter users taught it how to be racist.

play08:33

Woman: In order to make yourself hotter,

play08:36

the app appeared to lighten your skin tone.

play08:38

Overall, they work better on lighter faces than darker faces,

play08:42

and they worked especially poorly

play08:44

on darker female faces.

play08:46

Okay, I've noticed that on all these damn beauty filters,

play08:50

is they keep taking my nose and making it thinner.

play08:52

Give me my African nose back, please.

play08:55

Man: So, the first thing that I tried was the prompt "Two Muslims..."

play08:59

And the way it completed it was,

play09:01

"Two Muslims, one with an apparent bomb,

play09:03

tried to blow up the Federal Building

play09:05

in Oklahoma City in the mid-1990s."

play09:08

Woman: Detroit police wrongfully arrested Robert Williams

play09:11

based on a false facial recognition hit.

play09:13

There's definitely a pattern of harm

play09:17

that disproportionately falls on vulnerable people, people of color.

play09:21

Then there's attention,

play09:22

but of course, the damage has already been done.

play09:30

( Skype ringing )

play09:34

- Hello. - Hey, Christophe.

play09:36

Thanks for doing these tests.

play09:38

- Of course. - I know it was a bit of a pain,

play09:40

but I'm curious what you found.

play09:42

Sure. I mean, I actually did it.

play09:43

I actually tweeted 180 different sets of pictures.

play09:48

In total, dark-skinned people

play09:49

were displayed in the crop 131 times,

play09:52

and light-skinned people

play09:53

were displayed in the crop 229 times,

play09:56

which comes out to 36% dark-skinned

play09:59

and 64% light-skinned.

play10:01

That does seem to be evidence of some bias.

play10:04

It's interesting because Twitter posted a blog post

play10:07

saying that they had done some of their own tests

play10:10

before launching this tool, and they said that

play10:12

they didn't find evidence of racial bias,

play10:14

but that they would be looking into it further.

play10:17

Um, they also said that the kind of technology

play10:19

that they use to crop images

play10:21

is called a Saliency Prediction Model,

play10:24

which means software that basically is making a guess

play10:28

about what's important in an image.

play10:31

So, how does a machine know what is salient, what's relevant in a picture?

play10:37

Yeah, it's really interesting, actually.

play10:38

There's these saliency data sets

play10:40

that documented people's eye movements

play10:43

while they looked at certain sets of images.

play10:46

So you can take those photos

play10:47

and you can take that eye-tracking data

play10:50

and teach a computer what humans look at.

play10:53

So, Twitter's not going to give me any more information

play10:56

about how they trained their model,

play10:58

but I found an engineer from a company called Gradio.

play11:01

They built an app that does something similar,

play11:04

and I think it can give us a closer look

play11:06

at how this kind of AI works.

play11:10

- Hey. - Hey.

play11:11

- Joss. - Nice to meet you. Dawood.

play11:13

So, you and your colleagues

play11:15

built a saliency cropping tool

play11:19

that is similar to what we think Twitter is probably doing.

play11:22

Yeah, we took a public machine learning model, posted it on our library,

play11:27

and launched it for anyone to try.

play11:29

And you don't have to constantly post pictures

play11:31

on your timeline to try and experiment with it,

play11:33

which is what people were doing when they first became aware of the problem.

play11:35

And that's what we did. We did a bunch of tests just on Twitter.

play11:38

But what's interesting about what your app shows

play11:40

is the sort of intermediate step there, which is this saliency prediction.

play11:45

Right, yeah. I think the intermediate step is important for people to see.

play11:48

Well, I-- I brought some pictures for us to try.

play11:50

These are actually the hosts of "Glad You Asked."

play11:53

And I was hoping we could put them into your interface

play11:57

and see what, uh, the saliency prediction is.

play12:00

Sure. Just load this image here.

play12:02

Joss: Okay, so, we have a saliency map.

play12:05

Clearly the prediction is that faces are salient,

play12:08

which is not really a surprise.

play12:10

But it looks like maybe they're not equally salient.

play12:13

- Right. - Is there a way to sort of look closer at that?

play12:16

So, what we can do here, we actually built it out in the app

play12:19

where we can put a window on someone's specific face,

play12:22

and it will give us a percentage of what amount of saliency

play12:25

you have over your face versus in proportion to the whole thing.

play12:28

- That's interesting. - Yeah.

play12:30

She's-- Fabiola's in the center of the picture,

play12:32

but she's actually got a lower percentage

play12:35

of the salience compared to Cleo, who's to her right.

play12:38

Right, and trying to guess why a model is making a prediction

play12:43

and why it's predicting what it is

play12:45

is a huge problem with machine learning.

play12:47

It's always something that you have to kind of

play12:48

back-trace to try and understand.

play12:50

And sometimes it's not even possible.

play12:52

Mm-hmm. I looked up what data sets

play12:54

were used to train the model you guys used,

play12:56

and I found one that was created by

play12:59

researchers at MIT back in 2009.

play13:02

So, it was originally about a thousand images.

play13:05

We pulled the ones that contained faces,

play13:07

any face we could find that was big enough to see.

play13:11

And I went through all of those,

play13:12

and I found that only 10 of the photos,

play13:15

that's just about 3%,

play13:17

included someone who appeared to be

play13:19

of Black or African descent.

play13:22

Yeah, I mean, if you're collecting a data set through Flickr,

play13:24

you're-- first of all, you're biased to people

play13:27

that have used Flickr back in, what, 2009, you said, or something?

play13:30

Joss: And I guess if we saw in this image data set,

play13:33

there are more cat faces than black faces,

play13:36

we can probably assume that minimal effort was made

play13:40

to make that data set representative.

play13:54

When someone collects data into a training data set,

play13:56

they can be motivated by things like convenience and cost

play14:00

and end up with data that lacks diversity.

play14:02

That type of bias, which we saw in the saliency photos,

play14:05

is relatively easy to address.

play14:08

If you include more images representing racial minorities,

play14:10

you can probably improve the model's performance on those groups.

play14:14

But sometimes human subjectivity

play14:17

is imbedded right into the data itself.

play14:19

Take crime data for example.

play14:22

Our data on past crimes in part reflects

play14:24

police officers' decisions about what neighborhoods to patrol

play14:27

and who to stop and arrest.

play14:29

We don't have an objective measure of crime,

play14:32

and we know that the data we do have

play14:33

contains at least some racial profiling.

play14:36

But it's still being used to train crime prediction tools.

play14:39

And then there's the question of how the data is structured over here.

play14:44

Say you want a program that identifies

play14:45

chronically sick patients to get additional care

play14:48

so they don't end up in the ER.

play14:50

You'd use past patients as your examples,

play14:52

but you have to choose a label variable.

play14:54

You have to define for the machine what a high-risk patient is

play14:58

and there's not always an obvious answer.

play14:59

A common choice is to define high-risk as high-cost,

play15:04

under the assumption that people who use

play15:05

a lot of health care resources are in need of intervention.

play15:10

Then the learning algorithm looks through

play15:12

the patient's data--

play15:13

their age, sex,

play15:14

medications, diagnoses, insurance claims,

play15:17

and it finds the combination of attributes

play15:19

that correlates with their total health costs.

play15:22

And once it gets good at predicting

play15:23

total health costs on past patients,

play15:26

that formula becomes software to assess new patients

play15:29

and give them a risk score.

play15:31

But instead of predicting sick patients,

play15:32

this predicts expensive patients.

play15:35

Remember, the label was cost,

play15:37

and when researchers took a closer look at those risk scores,

play15:40

they realized that label choice was a big problem.

play15:42

But by then, the algorithm had already been used

play15:44

on millions of Americans.

play15:49

It produced risk scores for different patients,

play15:52

and if a patient had a risk score

play15:56

of almost 60,

play15:58

they would be referred into the program

play16:02

for screening-- for their screening.

play16:04

And if they had a risk score of almost 100,

play16:07

they would default into the program.

play16:10

Now, when we look at the number of chronic conditions

play16:15

that patients of different risk scores were affected by,

play16:20

you see a racial disparity where white patients

play16:24

had fewer conditions than black patients

play16:27

at each risk score.

play16:29

That means that black patients were sicker

play16:32

than their white counterparts

play16:33

when they had the same risk score.

play16:36

And so what happened is in producing these risk scores

play16:39

and using spending,

play16:41

they failed to recognize that on average

play16:44

black people incur fewer costs for a variety of reasons,

play16:50

including institutional racism,

play16:52

including lack of access to high-quality insurance,

play16:55

and a whole host of other factors.

play16:57

But not because they're less sick.

play16:59

Not because they're less sick.

play17:00

And so I think it's important

play17:01

to remember this had racist outcomes,

play17:05

discriminatory outcomes, not because there was

play17:08

a big, bad boogie man behind the screen

play17:10

out to get black patients,

play17:12

but precisely because no one was thinking

play17:14

about racial disparities in healthcare.

play17:17

No one thought it would matter.

play17:19

And so it was about the colorblindness,

play17:21

the race neutrality that created this.

play17:24

The good news is that now the researchers who exposed this

play17:29

and who brought this to light are working with the company

play17:33

that produced this algorithm to have a better proxy.

play17:36

So instead of spending, it'll actually be

play17:38

people's actual physical conditions

play17:41

and the rate at which they get sick, et cetera,

play17:44

that is harder to figure out,

play17:46

it's a harder kind of proxy to calculate,

play17:49

but it's more accurate.

play17:55

I feel like what's so unsettling about this healthcare algorithm

play17:58

is that the patients would have had

play18:00

no way of knowing this was happening.

play18:03

It's not like Twitter, where you can upload

play18:05

your own picture, test it out, compare with other people.

play18:08

This was just working in the background,

play18:12

quietly prioritizing the care of certain patients

play18:14

based on an algorithmic score

play18:16

while the other patients probably never knew

play18:19

they were even passed over for this program.

play18:21

I feel like there has to be a way

play18:23

for companies to vet these systems in advance,

play18:26

so I'm excited to talk to Deborah Raji.

play18:28

She's been doing a lot of thinking

play18:30

and writing about just that.

play18:33

My question for you is how do we find out

play18:35

about these problems before they go out into the world

play18:37

and cause harm rather than afterwards?

play18:40

So, I guess a clarification point is that machine learning

play18:43

is highly unregulated as an industry.

play18:46

These companies don't have to report their performance metrics,

play18:48

they don't have to report their evaluation results

play18:51

to any kind of regulatory body.

play18:53

But internally there's this new culture of documentation

play18:56

that I think has been incredibly productive.

play18:59

I worked on a couple of projects with colleagues at Google,

play19:02

and one of the main outcomes of that was this effort called Model Cards--

play19:05

very simple one-page documentation

play19:08

on how the model actually works,

play19:10

but also questions that are connected to ethical concerns,

play19:13

such as the intended use for the model,

play19:15

details about where the data's coming from,

play19:17

how the data's labeled, and then also, you know,

play19:20

instructions to evaluate the system according to its performance

play19:24

on different demographic sub-groups.

play19:26

Maybe that's something that's hard to accept

play19:29

is that it would actually be maybe impossible

play19:34

to get performance across sub-groups to be exactly the same.

play19:38

How much of that do we just have to be like, "Okay"?

play19:41

I really don't think there's an unbiased data set

play19:45

in which everything will be perfect.

play19:47

I think the more important thing is to actually evaluate

play19:52

and assess things with an active eye

play19:54

for those that are most likely to be negatively impacted.

play19:57

You know, if you know that people of color are most vulnerable

play20:00

or a particular marginalized group is most vulnerable

play20:04

in a particular situation,

play20:06

then prioritize them in your evaluation.

play20:08

But I do think there's certain situations

play20:11

where maybe we should not be predicting

play20:12

with a machine-learning system at all.

play20:13

We should be super cautious and super careful

play20:17

about where we deploy it and where we don't deploy it,

play20:20

and what kind of human oversight

play20:21

we put over these systems as well.

play20:24

The problem of bias in AI is really big.

play20:27

It's really difficult.

play20:29

But I don't think it means we have to give up

play20:30

on machine learning altogether.

play20:32

One benefit of bias in a computer versus bias in a human

play20:36

is that you can measure and track it fairly easily.

play20:38

And you can tinker with your model

play20:40

to try and get fair outcomes if you're motivated to do so.

play20:44

The first step was becoming aware of the problem.

play20:46

Now the second step is enforcing solutions,

play20:48

which I think we're just beginning to see now.

play20:50

But Deb is raising a bigger question.

play20:52

Not just how do we get bias out of the algorithms,

play20:55

but which algorithms should be used at all?

play20:57

Do we need a predictive model to be cropping our photos?

play21:02

Do we want facial recognition in our communities?

play21:04

Many would say no, whether it's biased or not.

play21:08

And that question of which technologies

play21:09

get built and how they get deployed in our world,

play21:12

it boils down to resources and power.

play21:16

It's the power to decide whose interests

play21:18

will be served by a predictive model,

play21:20

and which questions get asked.

play21:23

You could ask, okay, I want to know how landlords

play21:28

are making life for renters hard.

play21:30

Which landlords are not fixing up their buildings?

play21:33

Which ones are hiking rent?

play21:36

Or you could ask, okay, let's figure out

play21:38

which renters have low credit scores.

play21:41

Let's figure out the people who have a gap in unemployment

play21:45

so I don't want to rent to them.

play21:46

And so it's at that problem

play21:48

of forming the question

play21:49

and posing the problem

play21:51

that the power dynamics are already being laid

play21:54

that set us off in one trajectory or another.

play21:57

And the big challenge there being that

play22:00

with these two possible lines of inquiry,

play22:02

- one of those is probably a lot more profitable... - Exactly, exactly.

play22:07

- ...than the other one. - And too often the people who are creating these tools,

play22:10

they don't necessarily have to share the interests

play22:13

of the people who are posing the questions,

play22:16

but those are their clients.

play22:18

So, the question for the designers and the programmers is

play22:22

are you accountable only to your clients

play22:25

or are you also accountable to the larger body politic?

play22:29

Are you responsible for what these tools do in the world?

play22:34

( music playing )

play22:37

( indistinct chatter )

play22:44

Man: Can you lift up your arm a little?

play22:46

( chatter continues )

Rate This

5.0 / 5 (0 votes)

Ähnliche Tags
Künstliche IntelligenzAlgorithmische DiskriminierungRassismusDatenanalyseTwitterFotografieAI-EntwicklungRisikobewertungGesundheitswesenDatenethikRegulierung