ETHICAL CONCERNS

IN
MACHINE LEARNING
SOC 2101 – Society, Environment and Engineering Ethics
Course teacher: Minhajul Bashir

9/11/2021 [email protected] 1
IMAGINE …
You are a Laravel* developer. Whenever you face any problem during your
development, you Google your problem, and try to get the best solution. You usually
get the best helps from the forum named Laracast, although there is another
wonderful forum named StackOverflow.
If you search your Laravel problem on some random day, which website is the most
likely to pop up as the first result?

9/11/2021 [email protected] 2
GOOGLE IS LEARNING
FROM OUR SEARCH
HISTORY!!!

9/11/2021 [email protected] 3
MACHINE LEARNING
Teaching computers how to learn from data to make decisions or predictions
The computer must be able to learn to identify patterns without being explicitly
programmed to
Part of the study of Artificial Intelligence

9/11/2021 [email protected] 4
MACHINE LEARNING
WORKFLOW

9/11/2021 [email protected] 5
WHAT DOES FACEBOOK
SHOW YOU IN YOUR
HOMEPAGE?
To be more specific, stories of which friends do you see?

9/11/2021 [email protected] 6
FACEBOOK’S MACHINE
LEARNING
Facebook’s policy is to show you contents of those friends who you interact with the
most
🢝 Reactions
🢝 Comments
🢝 Chats
🢝 Post/comment/photo tags

If you reduce communicating with your (once) best friends, Facebook will probably
make you forget them
If you are fond of memes, Facebook will show you more memes
If you are fond of propaganda news, Facebook will show you the same

9/11/2021 [email protected] 7
FACEBOOK IS LEARNING
FROM OUR
INTERACTION WITH
EVERY CONTENT!!!

9/11/2021 [email protected] 8
ETHICAL CONCERNS IN
MACHINE LEARNING
Is the target system ethical?
Is the process ethical?
Is the data source ethical?
Is the data ethical?
Is the impact ethical?

9/11/2021 [email protected] 9
PRIVACY AND SURVEILLANCE
Access to private data and personally identifiable data
Aspects of privacy –
🢝 The right to be left alone
🢝 Information privacy
🢝 Control over information about oneself
🢝 The right to secrecy

9/11/2021 [email protected] 10
PRIVACY AND SURVEILLANCE
Our modern digital lifestyle leads to digital collection of our data
Possibilities of intelligent data collection, and data analysis
“In this vast ocean of data, there is a frighteningly complete picture of us”

9/11/2021 [email protected] 11
PRIVACY-PRESERVING
TECHNIQUES
Anonymization
Access control
Encryption

9/11/2021 [email protected] 12
MANIPULATION OF BEHAVIOR
Information collected may be used to manipulate behavior, online and offline
User’s intense interaction with data systems make him/her vulnerable to “nudges”
especially designed for him/her
🢝 Make one unintentionally purchase something/subscribe to some service
🢝 Manipulate one’s thought process

9/11/2021 [email protected] 13
THE CAMBRIDGE
ANALYTICA SCANDAL
The power of data-driven behavior manipulation

9/11/2021 [email protected] 14
CAMBRIDGE ANALYTICA
A British political consulting firm which collected huge number of data from Facebook
🢝 used them for building psychographic profiles of the users

Cambridge Analytica was a consultant of Donald Trump’s 2016 US Election campaign, and
Leave.EU’s Brexit campaign!!

The campaign team used micro-targeting technique

🢝 Trump supporters were shown triumphant visuals of him
🢝 Swing voters were informed of Trump’s notable supporters, and negative ideas of his opponent
Hillary Clinton

Was Cambridge Analytica’s works ethical?

Let’s discuss
9/11/2021 [email protected] 15
BIAS IN DECISION SYSTEMS
Bias due to discriminatory preconception about members of a group
Cognitive biases
Statistical biases
Historical biases

9/11/2021 [email protected] 16
AMAZON’S BIASED
RECRUITING TOOL

9/11/2021 [email protected] 17
AMAZON’S AUTOMATING
RECRUITING PROCESS
AI reviewed job applicant’s resumes and rated applicants so that recruiters don’t
spend much time on resume screening.
Amazon had used the last 10 years’ historical data to train their AI model. The data
was male dominated as 60% of Amazon’s employees consisted of men.

What kind of bias is likely to be introduced here?

9/11/2021 [email protected] 18
RACIAL BIAS IN
HEALTHCARE RISK
ALGORITHM

9/11/2021 [email protected] 19
RACIAL BIAS IN HEALTHCARE
RISK ALGORITHM
AI predicts which patient is likely to need extra medical care. More than 200 million
U.S. citizens data was used.

What kind of bias is likely to be introduced here?

What was the reason for bias?

9/11/2021 [email protected] 20
HOW TO FIX BIASES IN ML
ALGORITHMS?
Remove data bias
🢝 not fully possible

Remove algorithmic bias

🢝 Should we remove columns likely to cause bias (such as sex or race) ?

ML interpretability
🢝 understand why a model is making a particular decision

9/11/2021 [email protected] 21
OPACITY OF AI SYSTEMS
AI systems will extract patterns from a given dataset
🢝 Many AI systems are black boxes (such as deep neural network, random forest)
🢝 Even programmers who designed the system do not know how the system identified a pattern

AI systems are thus opaque to the user and even to the experts usually
🢝 What if they learn the wrong thing? Difficult to trust its decision

We need explainable/interpretable ML in high-stakes decision makings (e.g. medical

diagnostic software)
🢝 to make the system fair and transparent, and thus trustworthy
🢝 to protect consumers

9/11/2021 [email protected] 22
REFERENCES
Bias in AI: What it is, Types & Examples of Bias & Tools to fix it:
https://research.aimultiple.com/ai-bias/

9/11/2021 [email protected] 23