Heads up! To view this whole video, sign in with your Courses account or enroll in your free 7-day trial. Sign In Enroll

Ethical Design

Well done!

You have completed Ethical Design!

Sign up for Treehouse Back to Library
Preview
Video Player
00:00
00:00
00:00
  1. 2x 2x
  2. 1.75x 1.75x
  3. 1.5x 1.5x
  4. 1.25x 1.25x
  5. 1.1x 1.1x
  6. 1x 1x
  7. 0.75x 0.75x
  8. 0.5x 0.5x
Use Up/Down Arrow keys to increase or decrease volume.

Perpetuating Bias and Amplifying Hate with Artificial Intelligence

6:52 with Hope Armstrong and Michelle Zohlman

Machine learning opens up a substantial ethical debate, as it can perpetuate bias and amplify hate. We'll look at real-world examples where AI got it wrong, and then we'll explore some best practices to prevent mistakes.

MICHELLE: You've probably heard about self driving cars, voice assistance, or 0:00
some artificial intelligence approaches to science and healthcare. 0:04
Like IBM Watson that helps us dig through vast libraries of complicated 0:08
information in an instant. 0:12
All of these things are powered by machine learning, 0:14
which is a subset of artificial intelligence. 0:18
Artificial Intelligence, abbreviated as AI, 0:21
is the overarching concept of creating intelligent 0:25
machines that can simulate human thinking. 0:30
Machine learning is a subset of AI and it is the ability of the computer to 0:35
analyze information and then create new rules. 0:40
So instead of humans doing all of the programming, 0:43
we can instead point computers in the right direction and 0:46
they can continue on a path of self improvement. 0:50
We're giving computers the chance to learn and create on their own. 0:53
Machine learning opens up a substantial ethical debate. 0:56
Cathy O'Neil, author of the book Weapons of Math Destruction, 1:02
explains it like this. 1:06
Big Data processes codify the past, they do not invent the future. 1:07
Doing that requires moral imagination and that's something only humans can provide. 1:14
We have to explicitly embed better values into 1:20
our algorithms creating Big Data models that follow our ethical lead. 1:23
Sometimes that will mean putting fairness ahead of profit. 1:27
Increasingly, machine learning is organizing information such as search 1:31
results according to invisible parameters that reveal biases in society. 1:37
Even though developers may not identify as racist or 1:42
sexist, the algorithms or the data fed to the system might be. 1:47
The software we create can amplify inequalities to make them even worse. 1:51
HOPE: The data that informs machine learning can have biases, 1:58
this is called algorithmic bias. 2:03
In 2016, University of Virginia computer science professor Vicente Ordonez, 2:06
noticed a sexist pattern in two image collections commonly used for research. 2:12
Images of shopping and washing are linked to women, 2:17
images of coaching and shooting are linked to men. 2:21
This is just one example of how machine learning can amplify sexism. 2:26
It can amplify racial bias too. 2:30
In a TEDx Talk by Joy Buolamwini, 2:34
she explains how facial recognition software didn't detect her face. 2:37
Those who develop the algorithm hadn't trained it to 2:43
recognize a range of skin tones and facial structures. 2:47
In addition to perpetuating biases, 2:51
technology can shape our perception of reality. 2:53
Guillaume Chaslot, a former software engineer at YouTube created 2:56
the algorithm that predicts which video to play next. 3:00
He admitted the algorithm tries to find rabbit holes that will draw 3:04
you into more videos on a given topic. 3:09
Algorithms optimized for 3:12
increasing watch time can inadvertently lead to radicalization. 3:14
A comprehensive study in Science Magazine found that fake news is 3:19
shared six times more than accurate stories. 3:23
When systems are not factually checked, 3:27
they use virality as the only success metric. 3:30
Interestingly, this is how conspiracy theories are spread. 3:33
Similarly, something as innocuous as 3:38
autocompletion in the Google search bar shapes our reality. 3:41
Let's search for "climate change is." 3:45
The search prediction makes recommendations based on 3:52
your search history as well as trending searches in your area. 3:55
I live in an area that's supportive of reversing climate change, so 3:58
my results are mostly reflective of that. 4:03
In a place resistant to climate change, the results may look very different. 4:06
While google.com generally looks the same for everyone, 4:11
it's subtle nudges like these that keep people in differing realities. 4:15
It's one of the reasons why political polarization is at a 30 year high. 4:20
Rashida Richardson is the Director of Policy Research at AI Now, 4:27
a research institute studying the social implications of artificial intelligence. 4:32
She has said that while there are anti-discrimination laws to protect 4:38
people regarding housing, credit, and employment, they have limits. 4:42
Racial disparities persist because of historical exclusion. 4:47
In a paper she authored, the team researched predictive policing 4:50
systems that forecast criminal activity. 4:55
The predictions are flawed because they're trained on racially discriminatory data. 4:59
The system took corrupt, illegal, and racially biased police 5:04
activity in the past and created more of that for the future. 5:10
MICHELLE: So what do you think? 5:15
Should computers be making automatic decisions? 5:17
Is it okay in some applications, but not in others? 5:19
Consider its application in finance, insurance, 5:23
healthcare, employment, and law enforcement. 5:27
Should algorithms amplify any content? 5:30
Who should decide? 5:33
As machine learning is still relatively new, 5:35
I don't think anyone has perfected this yet. 5:38
Here are a few ideas. 5:41
Carefully select training data. 5:42
Ensure the algorithm is fed a dataset that reflects a wide spectrum of humanity and 5:46
audit for biases. 5:52
Put measures in place to regulate the system once it's released to the world. 5:53
If an automated process relies on an algorithm to review it, 5:59
provide an option for a manual review. 6:03
An algorithm cannot understand the nuances of a complicated decision as a human can. 6:06
Sometimes AI is not an appropriate solution, it has limitations. 6:13
Build diverse teams. 6:17
A diverse group minimizes bias and products and 6:21
better serves a broad audience. 6:24
Do your due diligence before releasing a product by 6:27
conducting user testing with a diverse range of participants. 6:31
Al is an emerging technology that's just beginning to develop best practices. 6:36
Do you have any other ideas for making artificial intelligence ethical? 6:41
For more information, 6:46
check out the Algorithmic Justice League link in the teacher's notes. 6:47