1 of 75

Data Ethics: Choices and Values

1

original slides & content by Kathleen Creel, Diana Acosta-Navas

CMSC 320 - Introduction to Data Science,

2025

  • Fardina Fathmiul Alam

2 of 75

2

BIG THANKS: Zico Kolter (CMU)

& Amol Deshpande (UMD) for the diagram

3 of 75

Ethics

What is right, what is fair, what is just.

3

4 of 75

In an ideal world

4

Established by a Society; allowed by the law

May not explicitly required by the law; align with principles of fairness, honesty, compassion, and respect for human dignity

5 of 75

In the real world

5

6 of 75

We will talk about

DATA ETHICS

6

The set of principles and processes that guide the ethical collection, processing, analysis, use and application of data having an effect on human lives and society

A branch of ethics that evaluates data practices

Ref: d’Aquin et al, Towards an “Ethics in Design” methodology for AI research projects, in AIES 2018

7 of 75

We use data to inform our decisions

7

  • Evidence-based
  • Impartial
  • Reliable

In Data Science:

8 of 75

What can we learn from a data set?

  • Patterns
  • Correlations
  • Distributions

  • Values

8

9 of 75

What can we learn from a data set?

9

How to interrogate a data set to ethically find relevant elements?

Values in Design

Data is intrinsically values-laden

10 of 75

VALUES IN DESIGN

10

PROBLEM FORMULATION

DATA INTERPRETATION

DATA

COLLECTION

11 of 75

What are values?

value (n): an individual or community’s belief about what matters

Values express what we care about

Efficiency

Privacy

Truth

Security

Beauty

Fairness

Values reveal our assumptions about the world, people interacting with our designs, and how our choices affect them

11

12 of 75

Intentionality of values

Explicit values: Values that designers intend for their products to embody

Collateral values: Values that crop up as side effects of design decisions and the way users interact with those designs. These values are not intentionally designed into the system.

I.e: Social Media

12

13 of 75

Explicit Values

13

Contact-tracing

Health Safety Efficiency

Public interest

14 of 75

Collateral Values

  • Security?
    • Where is information stored?
    • Encryption?
  • Privacy?
    • Who has access to information?
    • Geolocation or bluetooth?
    • What information is accessible to health authorities/ the public?
  • Autonomy?
    • Informed consent?

14

https://www.theverge.com/2021/4/27/22405425/android-google-contact-tracing-bug-privacy

https://www.jdsupra.com/legalnews/class-action-filed-against-commonwealth-3421108/

15 of 75

“A better world won’t come about simply because we use data; data has its dark underside.”

― Mike Loukides, Ethics and Data Science

15

16 of 75

16

https://www.forbes.com/sites/kashmirhill/2012/02/16/how-target-figured-out-a-teen-girl-was-pregnant-before-her-father-did/?sh=48f798316668

Is it Legal?

Is it Ethical?

Remember, data can be powerful, yet it also has its challenges and ethical considerations. A important question that raised is how to respect individuals lives when using their data.

17 of 75

Why Ethics?

DALL-E 2 is an advanced AI capable of generating art, potentially raising questions about its impact on traditional artistic roles.

  • What impacts on human artists' livelihoods?
  • Preserving human creativity?
  • How to maintain the value of artistic expression?

https://onsitego.com/blog/dall-e-2-tested-how-close-is-this-ai-to-replacing-artists/

18 of 75

Why Ethics?

"In 2017, the Rohingya were killed, tortured...and displaced in the thousands as part of the Myanmar security forces’ campaign of ethnic cleansing. In the months and years leading up to the atrocities, Facebook’s algorithms were intensifying a storm of hatred against the Rohingya which contributed to real-world violence,” said Agnès Callamard, Amnesty International’s Secretary General."

https://www.amnesty.org/en/latest/news/2022/09/myanmar-facebooks-systems-promoted-violence-against-rohingya-meta-owes-reparations-new-report/

Case Study: Spread of harmful anti-Rohingya content

19 of 75

Major Areas of Concern in Data Ethics

  • Problem Formulation
  • Data Collection
  • Data ownership
  • Data Privacy
  • Data Anonymity
  • Data Validity
  • Algorithm, Statistics Fairness
  • And many more…

19

20 of 75

Data Ethics: Choices and Values

20

PROBLEM FORMULATION

LANGUAGE

BIAS AND REPRESENTATION

FAIRNESS

Why should we care? Who benefits?

Who’s harmed? What data?

21 of 75

Problem Formulation Statements

  • Formulating a problem means treating the desired solution as good or worthy of being done.
  • Why should we care about solving this problem?
  • Who can agree that this is a problem worth solving?
  • Who would benefit from its solution?

21

Ultimately the answers to these questions will depend on the values and assumptions of the people being asked.

22 of 75

Example: Unsubscribe feature

  1. People don’t like our app/service & it should be easy for them to unsubscribe

Potential design: Unsubscribe button on app homepage. Single click and user is unsubscribed.

  • The user doesn’t like our app/service, we want to understand why so we can make it better for future users

Potential design: Unsubscribe button on app homepage 🡪 10 minute questionnaire 🡪 talk with a representative

22

23 of 75

Problem Formulation Statements

23

What is the problem to be solved?

  • Is Professor X a good teacher?
  • Do students think she is a good teacher?
  • Do most students think she is a good teacher?

24 of 75

Choice of Data

24

  • What kind of data should inform our decisions?
  • Where will it come from?
  • Is it a reflection of what we want to measure?

25 of 75

DATA COLLECTION

25

26 of 75

Data Collection

How can you be sure that your data has been ethically sourced?

26

No data can be collected until explicit consent for that purpose has been given and that consent can be retracted at any time.”

Informed Consent and Transparency is important!

27 of 75

GDPR AND CONSENT

General Data Protection Regulation – new law in EU that recently went into play

Requires unambiguous consent

    • data subjects are provided with a clear explanation of the processing to which they are consenting
    • the consent mechanism is genuinely of a voluntary and "opt-in" nature
    • data subjects are permitted to withdraw their consent easily
    • the organisation does not rely on silence or inactivity to collect consent (e.g., pre‑ticked boxes do not constitute valid consent);

27

28 of 75

According to GDPR (General Data Protection Regulation)

“When collecting personal information, organisations must act transparently and with consent, collecting data only for the explicit purpose it’s needed.”

28

The GDPR is one of the most comprehensive and toughest data protection laws in the world and provides an overarching framework for the processing of personal data.Though it was drafted and passed by the European Union (EU), it imposes obligations onto organizations anywhere, so long as they target or collect data related to people in the EU. The regulation was put into effect on May 25, 2018.

There are more: California Consumer Privacy Act (CCPA): rights over personal information held by businesses; California Privacy Rights Act (CPRA); Health Insurance Portability and Accountability Act (HIPAA); Gramm-Leach-Bliley Act (GLBA): federal law that requires financial institutions to protect the privacy and security of consumers' personal financial information; Federal Trade Commission Act (FTCA); Children's Online Privacy Protection Act (COPPA) etc.

29 of 75

Case Study: It's Not OKCupid

The popular dating platform OkCupid was used in an experiment to determine if love is actually blind.

  • The executives of the site wanted to put their matching algorithm on trial, so they completed a test on random users.
  • The research concluded that if a user was told they had a high compatibility score with another user, they were more likely to reach out to them.
  • This raised ethical concerns about manipulating users' emotions and happiness without their explicit consent.

29

30 of 75

Data Ethics: Choices and Values

30

PROBLEM FORMULATION

LANGUAGE

BIAS AND REPRESENTATION

FAIRNESS

Descriptive: what is : Describes facts or observations without judgment.

Normative: what should be : Involves values or ethics about what is right or wrong.

Thick normative: both

how data can be framed or analyzed based on both observed facts and ethical considerations.

31 of 75

Language

31

32 of 75

Descriptive terms? Evaluative terms?

32

33 of 75

Descriptive vs. Normative Language

33

Descriptive language

  • Statements of fact
  • What people did
  • What happened
  • “Lectures are

90-minutes long”

  • “Assignments take more than two hours to finish”

  • “Sections are mandatory”

34 of 75

Attendance: Not Mandatory

25

Textbook Required

Example: Descriptive Terms

35 of 75

Descriptive vs. Normative Language

Normative language:

  • Evaluative statements
  • Express the speaker’s opinions/reactions
  • How they think things should be

35

  • “right”

  • “wrong”

  • “good”

  • “bad”

  • “should”

  • “should not”

36 of 75

AWESOME

27

GREAT TEACHER

Example: Normative Language

37 of 75

Descriptive language

Normative language:

  • Evaluative statements
  • Express the speaker’s opinions/reactions
  • How things should be
  • Statements of fact
  • What people did
  • What happened
  • How things “are”

28

38 of 75

38

CLEAR

EASY TO

LISTEN TO

38

Distinction between normative & descriptive language is not always clean-cut

39 of 75

Thick Normative Terms

Descriptive AND normative:

  • Thick normative terms express morally or aesthetically “loaded” descriptions

  • Cowardly
  • Cautious
  • Polite
  • Rude
  • Chill
  • Kind
  • Caring
  • Smart
  • Knowledgeable
  • Professional

39

40 of 75

Example: toxic speech classification & context

40

Bears suck!!!

41 of 75

Ex: AI tools’ “racy” score & gender bias

41

  • goal: identify “raciness,” or how sexually suggestive images are
  • images of women more racy than men
  • Use of thick normative term. Problematic?

Guardian. Feb 2023

42 of 75

Descriptive or Normative?

42

Does the program you are writing contain descriptive claims?

Do it contain normative claims or values?

How about thick normative terms?

43 of 75

Data Ethics: Choices and Values

43

PROBLEM FORMULATION

LANGUAGE

BIAS AND REPRESENTATION

FAIRNESS

Why should we care? Who benefits?

Who’s harmed? What data?

Descriptive: what is Normative: what should be Thick normative: both

44 of 75

What is Bias?

44

Statistical bias is the difference between measured results and “true” value.

This is the “neutral” or statistical meaning of the word bias. You will see it often in discussions of patterns in data.

45 of 75

Discrimination

45

  • Direct discrimination: discrimination resulting from a negative attitude toward the social group (e.g. animus or indifference)

  • Indirect discrimination: discrimination that does not result from such an attitude, but from rules and procedures constructed in a way that favors one group over another

46 of 75

Concern: Difference Type of Bias can

happen is data collection Step

Sampling Bias, Non Response Bias, Volunteer Bias, Social Desirability Bias, Measurement Bias, cultural bias or societal bias, many more ………

47 of 75

SAMPLING BIASES

Sampling effective at reducing the data you need to analyze

Ideally you want random sample

  • → Otherwise you need to account for bias, which can be tricky

Bias in sampling: need to be very careful when generalizing inferences drawn from a sample

→ Even for random samples

Questions to ask: How was the sample selected? Was it truly random? Potential biases? How were questions worded? How is missing data/attrition handled? Was the sample size large enough?

47

48 of 75

SOME POTENTIAL SOURCES OF BIASES

Sample Bias

    • Selection bias: some subjects more likely to be selected
    • Volunteer bias: people who volunteer are not representative
    • Nonresponse bias: people who decline to be interviewed

Survey/Response Bias

    • Interviewer bias
    • Acquiescence bias – tendency to agree with all questions
    • Social desirability bias: people are not going to admit to embarrassing things

And more …..

48

49 of 75

Case Study: Sample Bias in Twitter Users

Twitter provides rich data, including network, text, and sociodemographic information. Valuable for observing behavior and opinions over time.

Scenario: suppose you want to predict the outcome of a US presidential race with two candidates: a democrat and a republican. A clever data science project is set up, and suppose you are able to get an accurate reflection of the political preference for all Twitter users as being a democrat (Y = 0) or a republican (Y = 1) voter.

Reference: Textbook DM

50 of 75

Case Study: Sample Bias in Twitter Users

Twitter users may not represent the broader population.

(1) overrepresentation in densely populated areas.

(2) gender bias, which is decreasing over time.

(3) non-representative racial and ethnic distributions.

According to “Research Center Study (2019)"

  • Twitter users are younger.
  • More likely to identify as Democrats.
  • Highly educated.
  • Higher income compared to the overall U.S. population.

51 of 75

Reproducing social biases

57

BIAS AND REPRESENTATION

reproduces social biases:

(e.g. stereotypes)

often aligns most with dominant US culture

intellectual teacher

pleasant teacher

52 of 75

I have data about people! Now what?

52

Check for Statistical Bias

What correlations and patterns exist in my dataset?

In what ways do they fail to accurately represent the world?

Check for Discriminatory Bias

In what ways do the biases compound existing injustice?

Decide how to use the data given bias

  • For what social purposes would it be appropriate to use this data?
    • How should we communicate information about possible biases?

53 of 75

Data Ownership

Unlawfully acquire photo from other person’s mobile without permission or consent.

53

Is it ethically correct?

Are this person owner now?

Who owns Data? Who has legal rights and complete control to make decisions about the data they generate or possess?

54 of 75

Data Privacy

How to protection of individuals' personal information and the control (privacy right) they have over its collection, use, and disclosure?

54

Is it ethical to disclose your photo to unauthorized parties?

Is it ethical to use your photo without your consent?

55 of 75

Data Anonymization

The trouble with anonymization:

For every amount of data you might release, there exists some set of priors that, if possessed by an attacker, will let them reverse your anonymization.

56 of 75

A survey on the patient’s medical condition

56

Which Dataset is maintaining Anonymity ?

Dataset (A)

Dataset (B)

57 of 75

Data Validity

The integrity and accuracy of the data being collected, analyzed, and used for various purposes.

  • Involves ensuring that the data accurately represents the real-world phenomena or entities it is intended to measure or describe.
    • How well does data meet certain criteria
  • Valid data is reliable and free from errors, biases, or distortions that could lead to incorrect conclusions or decisions.

57

58 of 75

Data Validity: Is collected data valid?

58

Surgery is always performed in a hospital and if the data shows otherwise, the data is invalid.

Ref: Source

Hospital Data Rules

59 of 75

CASE STUDY: NETFLIX PRIZE

Raw data: 100M dated ratings from 480K users to 18K movies

Why released: improve predicting ratings of unlabeled examples

How anonymized: exact details not described by Netflix

    • All direct customer information removed
    • Only subset of full data; dates modified; some ratings deleted,
    • Movie title and year published in full

Attacks: dataset is claimed vulnerable [Narayanan Shmatikov 08]

    • Attack links data to IMDB where same users also rated movies
    • Find matches based on similar ratings or dates in both

Consequences: rich source of user data for researchers

    • unclear if attacks are a threat—no lawsuits or apologies yet

60 of 75

A company uses an algorithm to screen job applicants for interviews

How about Algorithmic Fairness?

60

Result

Potential Candidates

Name

Age

Skills

John

25

Nick

25

Bob

25

….

….

Is this outcome of algorithm ethically fair?

61 of 75

Misuse of Statistics / Misrepresenting Data

The unethical practice of distorting or manipulating statistical information to support a particular agenda, mislead others, or draw false conclusions.

61

Example: Selective collection of positive customer feedback as representative of overall customer satisfaction.

62 of 75

62

A statistical anomaly in the form of an excess just below 0.05

63 of 75

SCIENTIFIC METHOD: STATISTICAL ERRORS

Nature Article

P values not as reliable as many scientists assume

p-hacking: cherry picking data points etc., to get the p-values; repeating experiments if they fail till you get the result

Much discussion/debate about this issue in recent years

63

64 of 75

Social Issues

  • Society is pretty biased!
  • Your data is as well
  • Examples:
    • Crime statistics
    • Face recognition
    • Medical data
    • Judge verdicts

65 of 75

AI and Alignment

Most people working in AI think there is about a 5% chance that an AI will kill a large portion of humanity in the next 50 years or so.

I've heard that number go as high as 90% from people who work in AI alignment.

66 of 75

Always Remember

It is not always simple to decide what's right or wrong.

67 of 75

Trolley Problem!

67

It is not always simple to decide what's right or wrong.

Decide Whom to save?

This example shows how complex ethical questions can be, and how changing details can make the decision even harder. This relates to data ethics because making choices about data use can also involve difficult decisions where the right answer isn't always clear.

Data ethics is about navigating these complex situations, making thoughtful choices, and considering the impact on people's privacy, fairness, and well-being.

68 of 75

Book Recommended to read

Weapons of Math Destruction: How Big Data Increases Inequality and Threatens

Democracy by Cathy O'Neil

69 of 75

Data Science in Industry

(NOT NEEDED FOR THE COURSE)

70 of 75

WHAT IS A DATA SCIENTIST?

Many types of “data scientists” in industry …

  • Business analysts, renamed
    • “… someone who analyzes an organization or business domain (real or hypothetical) and documents its business or processes or systems, assessing the business model or its integration with technology.” – Wikipedia
  • Statisticians
  • Machine learning engineer
  • Backend tools developer

70

Thanks to: Zico Kolter

71 of 75

KEY DIFFERENCES

Classical statistics vs machine learning approaches

  • (Two are nearly mixed in most job calls you will see.)

Developing data science tools vs. doing data analysis

Working on a core business product vs more nebulous “identification of value” for the firm

  • → A data scientist working on a core business product might focus on developing machine learning models or algorithms to enhance the product's functionality.
  • → A data analyst tasked with identifying value for the firm may focus on analyzing market trends, customer behavior, and competitor performance to uncover strategic opportunities. They would conduct exploratory data analysis to understand key metrics, and more.

71

72 of 75

FINDING A JOB

Make a personal website.

  • Free hosting options: GitHub Pages, Google Sites
  • Pay for your own URL (but not the hosting).
  • Make a clean website, and make sure it renders on mobile:

Highlight relevant coursework, open source projects, tangible work experience, etc

Highlight tools that you know (not just programming languages, but also frameworks like Pytorch, TensorFlow and general tech skills)

72

73 of 75

“REQUIREMENTS”

Data science job postings – and, honestly, CS postings in general – often have completely nonsense requirements

  1. The group is filtering out some noise from the applicant pool
  2. Somebody wrote the posting and went buzzword crazy

In most cases (unless the position is a team lead, pure R&D, or a very senior role) you can work around requirements:

  • A good, simple website with good, clean projects can work wonders here …
  • Reach out and speak directly with team members
  • Alumni network, internship network, online forums

73

74 of 75

INTERVIEWING

We saw that there is no standard for being a “data scientist” – and there is also no standard interview style …

… but, generally, you’ll be asked about the five “chunks” we covered in this class, plus core CS stuff:

  • Software engineering questions: code efficiency, algorithm problem solving, version control etc.
  • Data collection and management questions (SQL, APIs, scraping, newer DB stuff like NoSQL, Graph DBs, etc)
  • General “how would you approach …” EDA questions
  • Machine learning questions (“general” best practices, but you should be able to describe DTs, RFs, SVM, basic neural nets, KNN, PCA, feature selection, clustering etc.)
  • Basic “best practices” for statistics, e.g., hypothesis testing, experimental design

Take-home data analysis project (YMMV): you may be given a take-home data analysis project as part of the interview process (vary company to company)

74

75 of 75

GRADUATE SCHOOL, ACADEMIA, R&D, …

Data science isn’t really an academic discipline by itself, but it comes up everywhere within and without CS

  • Modern science is built on a “CS and Statistics stack” …

Academic work in the area:

  • Outside of CS, using techniques from this class to help fundamental research in that field
  • Within CS, fundamental research in:
    • Machine learning
    • Statistics (non-pure theory)
    • Databases and data management
    • Incentives, game theory, mechanism design
  • Within CS, trying to automate data science (e.g., Google Cloud’s Predictive Analytics, “Automatic Statistician,” …)

75