Facebook conducted experiments on over 600,000 unwitting users to study how emotional content on the platform can spread and influence users' moods and posting behaviors. The experiments involved algorithmically reducing the amount of positive or negative posts users were shown in their news feeds for one week periods. Results found small effects, with exposure to fewer positive posts linked to slightly more negative language use and vice versa. Critics argue the research violated ethics guidelines by manipulating users without consent. Supporters counter that the risks were minimal and provided valuable insights to help Facebook understand and mitigate potential real harms from regular platform use.

2. Background: Facebook Practice
• Over 1 billion served/mo worldwide (in U.S., 128m
users/day)
• Average user: ~1500 items eligible for news feed
• Algorithm selects & shows ~300 items (to see all, go to
friend’s wall)
– Has always existed
– FB constantly changes it
– Proprietary
– Now based on ~100K criteria
– Almost certainly affects how much emotionally
charged content users see; very likely prioritizes 
posts
Matt McGee, “EdgeRank Is Dead: Facebook’s News Feed Algorithm Now Has Close To 100K
Weight Factors,” Marketing Land, Aug. 16, 2013,
http://marketingland.com/edgerank-is-dead-facebooks-news-feed-algorithm-now-has-close-to-100k-weight-55908

3. Three Hypotheses About
Facebook’s Practice
1. Social Comparison ( posts are risky)
• Small, observational studies
• Correlations b/w FB use & stress, jealously, loneliness, &
depression
• FB creates “self promotion-envy spiral”
1. Emotional Contagion ( posts are risky)
• Lab experiments & field data found happiness & depression
spread via in-person social networks
3. Null Hypothesis
•Krasnova, et al., Envy on Facebook: A Hidden Threat to Users’ Life Satisfaction?, (2013).
Wirtschaftsinformatik Proceedings 2013. Paper 92
•Kross, et al. (2013) Facebook Use Predicts Declines in Subjective Well-Being in Young Adults. PLoS
ONE 8(8): e69841
•Hatfield E, Cacioppo JT, Rapson RL (1993) Emotional contagion. Curr Dir Psychol Sci 2(3):96–100
•Fowler & Christakis (2008) Dynamic spread of happiness in a large social network: Longitudinal
analysis over 20 years in the Framingham Heart Study. BMJ 337:a2338
•Rosenquist, Fowler & Christakis (2011) Social network determinants of depression. Mol Psychiatry
16(3):273–281

4. Facebook’s PNAS Research
• 1 week (Jan. 11-18, 2012)
• Subjects “who viewed FB in English” randomly
selected by user ID
• 2 experiments manipulated extent to which subjects
(N = 689,003) were exposed to emotional content by
removing varying %’s of it from particular news feed
viewings
• All posts remained on friends’ walls; post not
presented in 1 viewing may appear in other viewings
• Software coded status updates as , , or neutral
(“no text was seen by the researchers”)
– Coded +/- if at least 1 +/- word, e.g., “nice” or “awful”

6. Facebook’s PNAS Research
Two parallel experiments; 4 conditions (for each, N = ~155,000)
Experiment 1: Removing  posts
•Treatment: Each  post “had between a 10% & 90% chance
(based on user ID) of being omitted…for that specific viewing.”
•Control: 10%–90% of 46.8% (i.e., 4.68%-42.12%) of eligible
posts randomly removed w/o regard to emotional content
Experiment 2: Removing  posts:
•Treatment: Each  post “had between a 10% & 90% chance
(based on user ID) of being omitted…for that specific viewing.”
•Control: 10%–90% of 22.4% (i.e., 2.24%-20.16%) of eligible
posts randomly removed w/o regard to emotional content

7. Results
Compared to control subjects, subjects exposed to fewer 
posts subsequently in their own posts:
•Used 0.1% fewer  words (Cohen’s d = 0.02)
•Used 0.04% more  words (Cohen’s d = 0.001)
•Produced only 96.7% as many words overall
Compared to control subjects, subjects exposed to fewer 
posts subsequently in their own posts:
•Used 0.07% fewer  words (Cohen’s d = 0.02)
•Used 0.06% more  words (Cohen’s d = 0.008)
•Produced only 99.7% as many words overall
Issues of methodology & interpretation:
•Questionable instrument: LIWC 2007 not intended for lengthy text
•Questionable coding: “I’m not having a great day.” “Oh great.”
•Effect on word choice ≠ effect on mood (no giggling at a funeral)

8. Tiny Effect Sizes
Source: Lee Becker,
Effect Size Calculators,
http://www.uccs.edu/lbeck
er/effect-size.html#Cohen

9. Like, Really Tiny
Cohen’s d = 0.1 (FB effects: d = 0.001–0.02)
(Visualization tool doesn’t support ES smaller than 0.1!)
Source: Kristoffer Magnusson, Interpreting Cohen's d effect size: an
interactive visualization, http://rpsychologist.com/d3/cohend/

10. Objections
• No IRB review
• No informed consent (except ToS)
• No debriefing
• Knowingly psychologically harmed users
– EPIC: “purposefully messed with people’s minds”
– Guardian: “deliberately made people sad”
– Slate: “intentionally made thousands upon thousands
of people sad”

11. Were the PNAS Experiments
Legally Subject to IRB Review?
• U.S. federal & most state jurisdictions
– Facebook: No.
– Cornell: only if it checked box AND “engaged in research”
• Maryland (never enforced?; likely unconstitutional as applied
to “speech” research)
– “A person may not conduct research using a human
subject unless [he] conducts the research in accordance
w/the [Common Rule].” Who “conducted” research?
– If HSR: At least FB. If QA/QI: Neither.

12. Was Cornell “Engaged in Research”?
— Cornell affiliates (w/FB) designed research & wrote paper
— FB data scientist “performed research” & “analyzed data”
OHRP Guidance on Engagement of Institutions in HSR
(Oct. 16, 2008):
Engagement when, inter alia, researchers obtain:
•data about subjects through intervention or interaction
•identifiable private information about subjects
•subjects’ informed consent
NO engagement when, inter alia, researchers:
•obtain coded private info from another institution involved in
the research (that retains a link to individually identifying info)
and are unable to readily ascertain subjects’ ID
•author a paper, journal article, or presentation describing
HSR

13. Should Cornell’s IRB Have Reviewed It
Anyway, As Good Policy?
In applying this guidance, it is important to note that at
least one institution must be determined to be engaged
in any non-exempt human subjects research project
that is conducted or supported by HHS (45 CFR
46.101(a)).

14. What If an IRB Had Reviewed It?
Michelle N. Meyer, John Lantos, Alex John London,
Amy L. McGuire, Udo Schuklenk, Lance Stell
(July 16, 2014)
28 additional ethicist signatories:
http://www.michellenmeyer.com/

15. The Common Rule and the
foundational tenets of research
ethics on which it’s based
require informed consent for all
human subjects research.

16. Why Isn’t Informed Consent Always
Legally Or Ethically Required for
Human Subjects Research?
Because research ethics is informed by
more than the principle of respect for
persons’ autonomy.
It’s also informed by beneficence &
justice.
Balancing these principles yields
exceptions to informed consent.

17. Beauchamp, staff philosopher
to Nat’l Comm. (1978)
•Respect for persons’
autonomy
•Beneficence
•Justice
Beauchamp & Childress (1977)
•Respect for persons’
autonomy
•Beneficence/Nonmaleficence
•Justice

18. Belmont on Beneficence &
Risk-Benefit Analysis
“Persons are treated in an ethical manner not only by
respecting their decisions and protecting them from harm, but
also by making efforts to secure their well-being.”
“Beneficence . . . requires that we protect against risk of harm
to subjects & also that we be concerned about the loss of the
substantial benefits that might be gained from research.”
“[B]eneficence often occupies a well-defined justifying role in
many areas of research involving human subjects.
“Research also makes it possible to avoid the harm that may
result from the application of previously accepted routine
practices that on closer investigation turn out to be
dangerous.”
“[E]stimates of the probability of harm [must be] reasonable,
as judged by known facts or other available studies/”

19. Belmont on Justice & Subject
Selection
“[S]election of research subjects needs to be scrutinized in
order to determine whether some classes . . . are being
systematically selected simply because of their easy
availability, their compromised position, or their
manipulability, rather than for reasons directly related to the
problem being studied.”
•The distribution of the risks and potential benefits of
research matters. Are research risks borne solely by
subjects, with the potential benefits of the research enjoyed
by others?
•Or are subjects drawn from the population that stands to
benefit from the research?

20. Belmont on Respect for Persons’
Autonomy & Informed Consent
Don’t “withhold information necessary to make a considered
judgment, when there are no compelling reasons to do so.”
“In most cases of research involving human subjects, respect for
persons demands that subjects enter into the research voluntarily
and with adequate information. In some situations, however,
application of the principle is not obvious.”
“A special problem of consent arises where informing subjects of
some pertinent aspect of the research is likely to impair the validity
of the research. In many cases, it is sufficient to indicate to subjects
that they are being invited to participate in research of which some
features will not be revealed until the research is concluded. In all
cases of research involving incomplete disclosure, such research is
justified only if it is clear that (1) incomplete disclosure is truly
necessary to accomplish the goals of the research, (2) there are no
undisclosed risks to subjects that are more than minimal, and (3)
there is an adequate plan for debriefing subjects, when appropriate,
and for dissemination of research results to them.”

21. When Isn’t Informed Consent
Legally Or Ethically Required for
Human Subjects Research?
(1) When minimal risk research can’t otherwise be
conducted. [Next slide]
or
(2) When an activity is designed to assure or
improve quality rather than contribute to
generalizable knowledge, in which case it doesn’t
meet the Common Rule’s definition of “research,”
(45 C.F.R. § 46.102(d)) [Data-guided practice & Shifting
perspective slides]
The Facebook experiments fit (1) & could easily
have been repackaged to fit (2).

22. §46.116(d): Waiver/Alteration of IC
1. No more than “minimal risk”
2. Won’t “adversely affect the rights & welfare of subjects”
3. Couldn’t “practicably be carried out” w/o
waiver/alteration
4. Debrief “whenever appropriate”
Minimal risk: incremental risks of research (§46.111(a)(2)) not
greater than those “ordinarily encountered in daily life or during
the performance of routine physical or psychological
examinations or tests” (§46.102(i))
• Risk: psychological harm from exposure to emotionally
charged text [for  posts,  posts, neither, or both? Is
this minimal risk or no foreseeable risk?]
• How did control & treatment arms compare to user
feeds controlled by FB’s algorithm in practice in Jan.
2012?
• To routine risk of emotionally charged text in other
contexts (newspapers, Twitter, blog comment section)?

23. Knowledge-Producing Activities Designed
to Enable Internal Data-Guided Practice
Practice
• Can include
innovation &
adaptation, but
it’s ad hoc and
intuition-driven
Human Subjects
Research
Data-Guided
Practice
• Systematic, data-guided
practice innovation &
adaptation (e.g., QA/QI,
CER, standard-of-care
research, learning
healthcare systems, A/B
testing)
• Designed to bring about
immediate improvements in
practice in particular
settings
• Research & practice are
integrated
• Participation not always
optional (consent to
minimal risk QI part of
consent to underlying
practice)
• Classically, distinct
activity occurring in
isolation
• Designed to
contribute to
generalizable
knowledge
• Participation usually
optional
• Often IRB review

24. Rethinking the Research/Practice
Distinction to Enable Data-Guided
Practices
Remember Tom Beauchamp, the drafter of the Belmont Report?
Congress charged the Nat’l Comm. with defining the boundary b/w
“biomedical and behavioral research and the accepted and routine
practice of medicine.” Why? Because research was assumed to be
dangerous while practice was assumed to be safe. They needed to be
distinguished so that research could be singled out for heightened
regulation. A lot turns on this distinction.
Tom now thinks it was
dangerously mistaken.
So do I. It has led us to
overprotect research
participants &
underprotect practice
participants.
Google this  talk on
YouTube & watch it. It’s
important.

25. Shifting
Perspectives
As we evaluate FB’s
research in the
foreground, let’s do so in
light of FB’s practice in
the background.
First, let’s stipulate that
people might be harmed
by exposure to
emotionally charged
social media text. (If you
don’t take that seriously,
why are you upset about
an experiment that very
modestly altered the
amount of such text users
saw for 1 week?)

26. Now Recall the Three Hypotheses
About Facebook’s Practice
1. Social Comparison ( posts are risky)
1. Emotional Contagion ( posts are risky)
3. Null Hypothesis (don’t worry; it’s just noise)
How ought Facebook to respond to conflicting
data about the risks of its service?

27. Shifting
Perspectives &
Inverting
Criticisms
“The world is just the A of the
A/B test”
— Duncan Watts, Oct. 2014,
MIT CodeCon Conference
Facebook has been accused of
abusing its power by
experimenting, but the
alternative is to use its power to
set A without knowing A’s
effects on users.
When do companies have an
ethical duty to conduct an
experiment as part of QA/QI?

28. Shifting
Perspectives &
Inverting
Criticisms
What’s the real “experiment”
(in the sense of exposure to
(un)known risks)? Subjecting
~310,000 to A/B testing or 1
billion to A? Who are the real
guinea pigs?
To avoid badly biased results,
FB “withh[e]ld information
necessary [for users] to make
a considered judgment” about
whether to participate in the
experiment. The alternative
was not producing (& hence
withholding) data about FB’s
everyday risks.

29. Shifting
Perspectives &
Inverting
Criticisms
Some have invoked Kant to
criticize FB for treating subjects as
mere means to its corporate ends.
What do you call a company that
doesn’t investigate credible claims
that its service harms users?
When subjects bear little or no
incremental risk from research &
stand to benefit from its results, we
should take a cue from the
Belmont Report & the Common
Rule and stop fetishizing autonomy
& consent & start embracing
welfare & solidarity. We’re all in
these practices — social media,
health care — together. Let’s make
sure they’re safe & effective.