1 of 30

Data Ethics: �Choices and Values

Veronica A. Rivera, Ph.D.

McCoy Family Center for Ethics in Society - HAI

Original slides and content by Kathleen Creel, Diana Acosta-Navas, and Benjamin Xie

2 of 30

We use data to inform our decisions

Evidence-based
Impartial
Reliable

2

3 of 30

What can we learn from a data set?

Patterns
Correlations
Distributions
…

Values

3

4 of 30

4

5 of 30

Data is intrinsically values-laden

5

6 of 30

VALUES IN DESIGN

6

PROBLEM FORMULATION

DATA INTERPRETATION

DATA

COLLECTION

7 of 30

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

7

When we are ideating or building technology, or running a research study, our decisions encode values, whether we intend for them to do so or not

A value is an individual or community’s belief about what matters

They express what we care about. For example, someone designing a website might have some values they want to prioritize in the design, such as the website loading quickly for users; they might care about the security of the personal information users are providing via the website; they might care about the aesthetics of the website.

Values also reveal our assumptions about the world, how people interact…benefits/harms

Sometimes values are in tension with one another. You might care about the aesthetic on your website, at the expense of the website’s functionality.

8 of 30

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.

8

9 of 30

PROBLEM FORMULATION

9

10 of 30

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?

10

11 of 30

Example: Unsubscribe feature

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

11

Suppose you are on the product design team for a company that operates an app on a subscription model. You’re trying to design an unsubscribe feature for the app.

[CLICK] One formulation of the problem is that you want to make it as easy as possible…

[CLICK] Another formulation of the problem is that you as a company want to understand why users don’t like the app or service…

Which problem you focus on leads to different solutions and therefore different user experiences

If you focus on the first, prioritize ease of use for user
[CLICK] Potential design
If you focus on the second, prioritize information and company growth
[CLICK] potential design…users end up unhappy

Ultimately you will have accomplished your task of designing an unsubscribe feature with whatever formulation you pick. But the outcome on users is very different.
[PAUSE]
Here these two problem formulations are in tension with one another, but they don’t always have to be, as we’ll see in the next example

12 of 30

Example: Course evaluations

Students have different learning needs. We want to help students decide which courses best fit their learning style�
We want to promote and hire good teachers

We want teachers to have the info they need to improve their classes

12

Now imagine you’re a university administrator trying to design an effective course evaluation form. There are three problem formulations, or things you care about: [read]

Unlike the subscription example, you can probably solve all of these with a single course eval form design. For each problem formulation, you might vary the questions that get asked and which stakeholders – students, teachers, higher level admin, the public – you make the responses available to.

To help students identify courses, you might include a question that asks current students in the class what they would like to say about the class to potential future students. Stanford does this and makes these responses viewable by students and instructors on carta.
You might also include information about the quality of an instructor
Or about students’ favorite and least favorite topic in the course

Even though we’re able to design for each of the three problem formulations, there’s still a challenge here:

Adding more questions = longer survey = fewer students answering it
So maybe a tradeoff we’re making is that we answer each of these problems a little less well than we would if we had taken a different approach. That also encodes particular values.

13 of 30

DATA COLLECTION

13

14 of 30

What is data bias: Two definitions

Statistical: Difference between measured and “true” value
Social: Human-created biases, such as stereotypes, that arise through embedding of values.

14

15 of 30

Examples of social biases in data

Confirmation bias

When we favor information that confirms or strengthens our beliefs or values

15

Selection bias

When the selection of study participants or data is not randomized, so the sample is not representative of the entire population intended to be analyzed

Survivorship bias

A form of selection bias in which ”winners” are overly focused on in a sample

Source: https://www.statice.ai/post/data-bias-types

16 of 30

Example: confirmation & selection bias

Confirmation bias

Selection bias

16

[CLICK]. Perhaps I think carbs for you and I want to prove it.

I obtain a dataset with relevant variables and regress death rate against carb intake and find an anti-correlation.

Then I regress obesity against carb intake and find a correlation. Now I really believe that carbs are bad for people so I go write a blog post about the second correlation, ignoring the first. Confirmation bias is dangerous in very complex situations like this one. What good for you means can vary widely.

If we only pay attention to information that supports our view, essentially just the tip of the iceberg, we miss the opportunity to understand what’s really going on.

Now let’s look at an example of selection bias: Perhaps I want to understand how important education is to people. So I conduct a study and only recruit stanford students because that’s access to. Turns out stanford students probably have different views on the importance of education relative to the broader population. Cherry picking data points in this way leads to an unrepresentative sample and wrong conclusions.

17 of 30

DATA INTERPRETATION

17

18 of 30

Descriptive and normative terms

18

19 of 30

Descriptive vs. Normative Language

Descriptive language

Statements of fact
What people did
What happened

19

“Lectures are 90-minutes long”

“Assignments take more than two hours to finish”

“Sections are mandatory”

20 of 30

20

Attendance:

Not

Mandatory

Textbook

Required

21 of 30

Descriptive vs. Normative Language

Normative language:

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

21

“right”

“wrong”

“good”

“bad”

“should”

“should not”

22 of 30

22

AWESOME

GREAT

TEACHER

23 of 30

Normative language:

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

Descriptive language

Statements of fact
What people did
What happened
How things “are”

23

24 of 30

24

CLEAR

EASY TO

LISTEN TO

24

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

25 of 30

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

25

We call these terms that straddle descriptive and normative terms, thick normative terms

The thing with these words is that we naturally treat them as descriptive: as stating the facts, when in reality they are colored by values.

When I say that someone is “professional,” in a course evaluation, that is not just a description of the person. It is an endorsement of them, which comes loaded with a lot of assumptions I make about what it means to be professional.

And so, what looks like an objective description is actually a reflection of my values and my assumptions, influenced by the specific context in which I was raised and educated and the things that were perceived as professional in that context.

But others might see my use of “professional” and interpret it as objective truth, or they might have a different notion of what professional means. This can both lead to biases as you’ll see in the hw, as well as misinterpretations of data, which we’ll see in the next example.

26 of 30

Example: toxic speech classification & context

26

Bears suck!!!

Imagine you’ve been hired by Reddit to build a state of the art toxicity classifier to detect hate speech on the site. Your classifer will be used across all subreddits.

- As your first step you need to rate a set of comments based on their level of toxicity

But your ratings are informed by your experiences and the contexts you’re aware of. For example, if you see the comment “bears suck” on a wildlife subreddit, you might not think much of it. Maybe your only notion of bears are the wild animal kind. So perhaps you give it a very low toxicity rating.
Now imagine you’re on the bayarea or stanford subreddit and you see these comments in reference to Berkeley students around the time of the Big Game. Maybe in that context comments like these should have higher toxicity ratings. Perhaps a Berkeley student reading this comment would feel hurt.
If you weren’t familiar with this use of the word “bears” in this particular context, you might have overlooked the fact that in one situation, “bear” refers to an animal; while in another, it might refer to a human

Additionally, the work “suck” is a thick normative term.

In the animal context, perhaps the poster meant to express annoyance at a bear for eating their camping food
In the human context, suck might be loaded with a much more negative connotation

27 of 30

We should not get rid of normative terms altogether

27

28 of 30

I want to collect & analyze data to solve a problem. What should I do?

28

Check for social biases in your experiment/research setup

Look for embedding of values in data during analysis

Work with multiple stakeholders/people to identify interesting problems

We’ve looked at various ways in which values are embedded within problem formulation, data collection, and data anlaysis, when trying to arrive at a data-driven solution to a problem.

By being aware that these values exist and how they might crop up across these three stages, you’re already better prepared to avoid some of them, or take mitigations, when working with data.

What are some concrete things you can do when trying to use data to solve problems?

When trying to identify problems to solve, it’s important to work with various stakeholders to capture what’s most interesting to multiple groups

When collecting data, make sure to check for social biases in your experiments or research setup

And finally, when analyzing data, especially textual data, make sure to look for potential values embedded in the data itself, such as normative or think normative terms. And to consider the potential downstream effects like we saw in the toxic speech example.

29 of 30

Explicit Values

29

Health

Safety

Efficiency

Public interest

Contact-tracing

Quick example,

In 2020, Apple and Google teamed up to develop a contact tracing app for iOS and Android devices to help detect the spread of COVID-19. Fun fact: a Stanford PhD student was involved in developing some of the cryptography for these apps.

These were put on people’s phones automatically as part of an update, rather than a user downloading the app. But people were still able to consent whether to use them or not.

Explicit values were clear at the time

Promote health and safety
Optimize detecting who might be sick. Efficiency could save lives
This was done in the public interest.

https://www.cnbc.com/2020/09/01/apple-google-will-build-coronavirus-contact-tracing-software-right-into-your-phone.html

https://www.apple.com/newsroom/2020/04/apple-and-google-partner-on-covid-19-contact-tracing-technology/

30 of 30

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?

30

There were also collateral values that arose out of moving quickly to get contract tracing apps deployed

For a few months there were some security issues. The Andriod version had an issue which leaked security info to all the other apps that were on your phone

Privacy concerns from users that these apps were a violation of individual privacy

Concerns around autonomy, about whether the consent was informed enough and misinformation that you did not need to consent

These were some collateral values -- we were trying to design for public interest and safety in a quick way but perhaps it compromised some other values that we also care about of security, privacy, and autonomy

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/

https://www.usatoday.com/story/news/factcheck/2020/06/14/fact-check-android-users-must-opt-into-covid-19-tracing-technology/5341250002/