A06 — Emotional Contagion: Facebook manipulated the emotions of 689,000 users

Category: Experiment without consent / research ethics / emotional manipulation Company/companies: Meta (Facebook), Cornell University, University of California San Francisco Years: January 2012 (experiment), June 2014 (publication and scandal) Status: Closed; article not retracted; Common Rule reform (2018) Card ID: A06

Metadata

Field	Value
Country/region	USA (experiment and publication); UK (ICO investigation)
Year revealed	June 2014 (after publication in PNAS)
Years of practice	One week in January 2012 (experiment)
Total penalty	No financial penalties
Currency	—
Legal basis	Common Rule (45 CFR 46), UK DPA 1998 (pre-GDPR), UCPD
Whistleblower/discoverer	Scientific publication in PNAS (unintended disclosure); Robinson Meyer (The Atlantic) — popularization
Number of affected	689,003 English-speaking Facebook users
Status (as of today)	Closed — Common Rule reform (2018); article still cited

TL;DR

In January 2012, for one week, Facebook manipulated the news feeds of 689,003 English-speaking users, deliberately showing them more emotionally positive or negative posts — to study the phenomenon of “emotional contagion” through social networks. None of the participants gave informed consent. The researchers and the company did not go through the standard Institutional Review Board (IRB) procedure — the ethics committee approval required for human subjects research. The results were published on June 17, 2014 in the prestigious journal PNAS (Proceedings of the National Academy of Sciences), in an article by Adam Kramer (Facebook), Jamie Guillory (UCSF), and Jeffrey Hancock (Cornell).

The publication initially went unnoticed, until on June 28, 2014 journalist Robinson Meyer from The Atlantic published an analysis titled “Everything We Know About Facebook’s Secret Mood Manipulation Experiment”. The article went viral within 24 hours. One of the largest research ethics crises of the internet era erupted. PNAS published a rare “Editorial Expression of Concern”. The UK ICO opened an investigation (closed without penalties). Cornell University argued that it did not need an IRB, because Hancock “received already-collected data” — an argument that the academic community considered strained. Jamie Guillory of UCSF went through IRB after the fact. Adam Kramer publicly apologized on his Facebook profile.

The case became the first major test of human subjects research ethics in the Big Tech era. It showed that “ToS consent” (Terms of Service) does not meet the standards of informed consent. It was an indirect driver of the Common Rule reform in the USA (2018) as well as an argument for strengthening the GDPR in the EU. For citizens — the first, drastic proof that we are part of continuous experiments on social platforms, and that platforms can deliberately alter our mood.

Timeline

January 2012 — seven-day experiment. 689,003 English-speaking Facebook users divided into a group with a reduced number of positive posts from friends (group A), a reduced number of negative posts (group B), and control groups.
2013 — data analysis by the Kramer-Guillory-Hancock team; article drafting.
January 14, 2014 — manuscript submission to PNAS.
April 2014 — acceptance for publication by PNAS. Peer review did not question the lack of IRB.
June 17, 2014 — online publication in PNAS (vol. 111, no. 24): Kramer, Guillory, Hancock, “Experimental evidence of massive-scale emotional contagion through social networks”.
June 26–27, 2014 — first mentions in academic blogs (James Grimmelmann, Katy Waldman in Slate).
June 28, 2014 — Robinson Meyer publishes the viral article in The Atlantic: “Everything We Know About Facebook’s Secret Mood Manipulation Experiment”. The case explodes in general media.
June 29, 2014 — Adam Kramer publishes a public apology on Facebook.
July 2, 2014 — Sheryl Sandberg (COO Facebook) in an interview with NDTV in India (during her New Delhi visit): “It was poorly communicated on our part.”
July 3, 2014 — PNAS publishes “Editorial Expression of Concern” (editor-in-chief Inder Verma); UK ICO opens an investigation.
September 23–24, 2014 — Douglas F. Gansler (Maryland Attorney General) announces analysis of complaints by Prof. James Grimmelmann regarding the study.
July 2014 — Cornell University issues a statement: Hancock “did not obtain access to identifying data”, therefore he did not need an IRB.
2014–2015 — extensive academic debate; including in Nature (“When is academic research ethical research?”, 2014), in PLOS ONE.
2015–2016 — the US Department of Health and Human Services (HHS) consults changes to the Common Rule.
January 2017 — publication of the final version of the “Revised Common Rule” (the “Final Rule”) in the Federal Register.
January 21, 2019 — Revised Common Rule enters into force, introducing, among other things, “broader consent” for secondary data.
2018 — the GDPR enters into force in the EU, explicitly requiring informed consent for processing for research purposes (art. 9).
To this day — the PNAS article remains unretracted, still cited (over 2,000 citations according to Google Scholar).

Mechanism

How the experiment worked

Research hypothesis: does exposure to other people’s emotions online cause “contagion” of emotions in the observer, just like in face-to-face contact?

Method:

689,003 English-speaking users of Facebook were randomly selected (all had to use the English-language interface; EU citizens who had the English interface were also included in the experiment).
They were divided into experimental groups:
- Group A (reduced positive) — the algorithm, with a probability of 10–90%, omitted positive posts from friends
- Group B (reduced negative) — the algorithm, with a probability of 10–90%, omitted negative posts from friends
- Control group — the normal algorithm (if the “normal algorithm” can be considered “natural”)
Post classification: the LIWC (Linguistic Inquiry and Word Count) algorithm labeled posts as “positive” or “negative” based on an emotional lexicon.
After a week, the participants’ own posts were analyzed — whether they became more positive/negative in their own statements.

Result: the effect was confirmed. Users exposed to fewer positive posts from friends wrote fewer positive posts themselves (and more negative ones), and vice versa. The statistical effect was small (Cohen’s d ≈ 0.001–0.02 depending on the variable; a very small effect, practically insignificant at the individual level), but statistically significant given such a large sample. It was precisely this “smallness of the effect” that was part of the controversy — critics asked whether it was worth manipulating the emotions of 700,000 people for such a marginal finding.

What the ethical problem was

Common Rule standards (USA, 45 CFR 46):

Informed consent of the participant — knowledge that one is in an experiment, conscious consent.
Debriefing after the experiment — explanation of what it consisted of.
Right to withdraw without consequences.
Parental consent for those under 18 years of age.

Facebook’s position: users gave consent by accepting the Terms of Service, which included a clause on using data for “improving our services, as well as conducting research”. This is defensible legally, but not ethically — the ToS is signed once, it has hundreds of pages, no one reads it. This is not informed consent in the classic sense.

Cornell’s position: Jeffrey Hancock’s university announced that the study did NOT require an IRB, because Hancock “received already-collected data” (Facebook conducted the experiment, Hancock analyzed the results). This interpretation is controversial. Critics (including James Grimmelmann, a law professor also at Cornell) argued that if the university knew the data would be produced thanks to an experiment that was not conducted ethically, then the university should have refused to participate.

UCSF’s position: Jamie Guillory of UCSF reportedly went through IRB only after the fact, when the media were already asking questions.

Discovery

How the disclosure happened

This was not a classic whistleblower disclosure — the article was published openly in a scientific journal, in accordance with all standards. The disclosure consisted in the fact that journalists and the public noticed what had actually been published.

Sequence:

Publication in PNAS — several days without notice
James Grimmelmann’s blog (Cornell law professor) “Illusion of Change” begins critical analysis
Robinson Meyer of The Atlantic conducts interviews with research ethicists and publishes the article on June 28, 2014
The article goes viral on social media (ironically — via Facebook, Twitter)
Within 48 hours — massive reaction from media and politicians

First publications

June 28, 2014 — Robinson Meyer, “Everything We Know About Facebook’s Secret Mood Manipulation Experiment”, The Atlantic
June 29, 2014 — Kashmir Hill, “Facebook Manipulated 689,003 Users’ Emotions For Science”, Forbes
June 30, 2014 — dozens of articles in The Guardian, WaPo, NYT, CNN

Key people

Study authors (central)

Adam D. I. Kramer — data scientist on the “Facebook Data Science Team” (at the time an openly academic team working with publications). Held a PhD in social psychology from the University of Oregon. First author of the article. Publicly apologized.
Jamie E. Guillory — postdoctoral researcher, University of California San Francisco (Health Communications team). Previously studied at Cornell.
Jeffrey T. Hancock — professor of communication at Cornell University (currently Stanford). Ironically: Hancock specializes in the psychology of online lying and communication ethics. His participation in the study provoked particularly sharp criticism — even experts in communication ethics take part in unethical projects.

Critics and commentators

James Grimmelmann — professor at Cornell Law School (!). The most detailed legal critique. He argued that the study could have violated the Common Rule despite the lack of federal funding, because Cornell as a federally registered institution (with “federalwide assurance”) is subject to the rules. The “Illusion of Change” blog became the primary source of criticism.
Michelle Meyer — at the time Harvard Law School, research ethics expert; defended the study from the perspective that A/B testing is commonplace in tech and that IRB standards should be reformed to reflect digital realities.
Clay Johnson — entrepreneur and policy researcher, wrote a cautionary tweet: “The Facebook ‘transmission of anger’ experiment is scandalous, terrifying, and disgusting.”
Senator Mark Warner — sent a letter to the FTC.
Maryland Attorney General Douglas Gansler — publicly criticized Facebook.

Journalists

Robinson Meyer (The Atlantic) — author of the viral article.
Kashmir Hill (then Forbes, today NYT) — one of the first to broadly popularize the topic; later specialized in privacy (author of a book on Clearview AI).
Vindu Goel (The New York Times) — detailed technical analyses.

Inside Facebook

Sheryl Sandberg (COO) — official statement: “It was a communication error.”
Mark Zuckerberg — did not speak publicly in 2014.

Regulators

Information Commissioner’s Office (UK) — opened an investigation on July 3, 2014; closed without penalty (argument of lack of jurisdiction over US-based Meta).
PNAS Editor-in-Chief Inder Verma — signed the “Editorial Expression of Concern”.
Office for Human Research Protections (OHRP) at HHS USA — initiated informal consultations, but not punitive proceedings.

Company response

Facebook

Adam Kramer — public apology (June 29, 2014) on his profile: “Our goal was never to upset anyone. I sympathize with anyone this caused distress to. In retrospect, the research benefits of the paper may not have justified all of this anxiety.”

Sheryl Sandberg (July 2, 2014, interview with NDTV in India, during her New Delhi visit): “The fact is that this study was part of ongoing research that companies conduct to test different products. It was poorly communicated on our part.”

Monika Bickert (Facebook Head of Policy Management) — publicly argued that Facebook always A/B tests: new design, new algorithms, new features. The PNAS study was “simply writing about it in an article”.

No formal promise to stop such experiments was ever made. Facebook still A/B tests everything today — it just doesn’t publish the results in PNAS.

Actions after the scandal:

Facebook established an internal review process for “sensitive” experiments (data scientists must obtain approval from an internal team).
Adam Kramer remained at Meta for years, though with a lower public profile.

Cornell University

Statement (June 30, 2014): Cornell did not approve or fund the study, and Hancock had access only to anonymized data after the fact. Therefore, an IRB was not required. This position was not modified despite criticism.

UCSF

The Guillory team went through a post-hoc IRB. UCSF did not issue a public declaration of distancing itself.

PNAS

July 3, 2014 — “Editorial Expression of Concern” — a rare move for such a prestigious journal. Verma wrote: “It is a matter of concern that the collection of the data by Facebook may have involved practices that were not fully consistent with the principles of obtaining informed consent and allowing participants to opt out.” The article was never retracted.

Legal proceedings

Jurisdictions

US federal — OHRP considered; no formal consequences
US state — Maryland AG public criticism; no proceedings
UK — ICO investigation, closed without penalties
EU (pre-GDPR) — directive 95/46/EC theoretically applied; no practical consequences

Legal basis

Common Rule (45 CFR 46) — federal requirements for research on humans. Facebook argued that it was not subject, because it had no federal funding. Cornell: “did not conduct the research”. UCSF: “conducted post-hoc IRB”.
45 CFR 46.116 (informed consent) — standard requirements violated.
UK Data Protection Act 1998 — legal basis for processing; Facebook argued that the ToS was sufficient.
Directive 95/46/EC — art. 7 (legal basis); unclear whether the ToS was sufficiently informed consent.

Key stages

Date	Stage
July 3, 2014	PNAS “Expression of Concern”; ICO opens case
2015–2016	Common Rule reform consultations
January 2017	Final Rule publication
January 2019	Common Rule reform enters into force

Penalties and settlements

Date	Body	Amount	Jurisdiction	Basis
—	—	No financial penalties	—	—

The only formal sanction: PNAS Editorial Expression of Concern (July 3, 2014). The article was not retracted.

Precedents and implications

For US law

Direct contribution to the Common Rule reform (2018). The new version introduces:
- “Broader consent” — a one-time consent to the use of data in future unspecified research (under certain conditions).
- Clarification of the scope of exclusions for research on secondary data.
- Better definition of “minimal risk research”.
Strengthening of the discussion on a federal privacy law, which nevertheless has not been enacted.

For EU law

Indirect impulse for GDPR art. 9(2)(j) — explicit consent for processing for research purposes.
Argument for “freely given” consent (recital 43 GDPR): a ToS with consent to “research” does not meet this requirement.

For Big Tech practice

Everyone A/B tests, but no one publishes in PNAS anymore. Big Tech has become significantly more cautious about publishing experimental results as scientific papers.
Internal ethics committees have become the norm at Meta, Google, Microsoft.
The reputational risk of emotional experiments is now widely recognized — Instagram after the Facebook Files announced that it is reducing A/B testing that alters algorithmic recommendations for teenagers.
OkCupid in 2014, following this case, published the post “We Experiment On Human Beings!” — defending the practice of A/B testing as an industry standard. This time the public reaction was milder.

Class actions

No significant class action lawsuits in the US. The UK ICO did not allow further steps.

Conclusions for citizens

Portal section — practical.

What does this mean for me?

Even if today, in 2026, Facebook doesn’t conduct experiments as overtly emotional as in 2012, you are still continuously a participant in hundreds of simultaneous A/B tests on all the social platforms you use. Meta, TikTok, YouTube, X, LinkedIn — each of them tests different variants of the recommendation algorithm, screen layout, colors, messages on you. You don’t know which group you’re in. You don’t know whether your mood today after scrolling is “normal for you” — or whether it’s the result of the algorithm testing a variant today that shows you more conflict-inducing content.

How to protect yourself?

Assume your feed is an experiment. Any day you may be in a different test group.
Diversify your sources of information. If one platform “tunes you depressively”, other sources can balance you out. RSS readers, email newsletters (outside the algorithm), radio, newspapers — old media have one advantage: they don’t experiment on you in real time.
Observe your mood after social media. If you notice that after 30 minutes on Instagram/Facebook you feel significantly worse than before — that’s a signal that the algorithm is currently testing a variant that doesn’t serve you.
Turn off algorithmic sorting where you can: Facebook “Most Recent”, X “Following”, Instagram “Following”.
Periodic cleaning of “follows”. Every 6 months, review who you follow — the algorithm filters what you see anyway, but at least you know what it can choose from.
If you are in depression or an emotional crisis state: Facebook/Instagram are the worst platforms in such a state. The algorithm may unintentionally deepen the decline. Take a break.

What rights do I have?

In the EU (GDPR):

Art. 21 — right to object to processing for direct marketing purposes (absolute)
Art. 22 — right not to be subject to automated decisions (including profiling) — limited in practice
Art. 9 — processing of sensitive data (including data on mental health status) requires explicit consent
GDPR art. 5(1)(a) — processing must be lawful, fair and transparent — an experiment without informing is prima facie inconsistent with this article

In the USA:

The Common Rule protects only academic research with federal funding
CCPA (California) gives a right to opt out of data sales, not for A/B tests

Where to report?

Poland: President of UODO
EU: Your national DPA
USA: FTC, OHRP (limited)

Note for mediators

The “Emotional Contagion” case has a direct implication for family and marital mediation: the mood of the parties before the session may be deliberately tuned by the algorithm. Before key mediation sessions, it’s worth suggesting to the parties a “digital detox” — 1–2 days without social media. This is not esotericism; it is reducing exposure to emotional experiments conducted by platforms. In business mediations, it’s worth asking whether the emotional escalation in the conflict occurred after an intense social media phase by one of the parties.

Context

A sample of 689,003, not “millions” — many popular descriptions incorrectly cite millions. But it’s still one of the largest social samples in the history of the social sciences. For comparison, Zimbardo’s classic Stanford Prison Experiment had 24 participants.
“Small effect, big PR” — the statistical effect was slight (Cohen’s d ≈ 0.001–0.02 depending on the variable; a very small effect, practically immeasurable at the individual level), but the social outrage enormous. It showed that even a small statistical emotional manipulation at a massive scale is sufficient to shock.
James Grimmelmann ironically worked at Cornell, the same university as Hancock. Grimmelmann was the sharpest critic of the study — which was described by Harvard Law Review as “Cornell kicking Cornell”.
Jeffrey Hancock after the scandal did not lose his position, later moved to Stanford. There he runs a laboratory on online trust research. The irony: a trust researcher participated in an experiment that undermined trust in platforms.
OkCupid 2014 — in reaction to the FB scandal, OkCupid published the post “We Experiment On Human Beings!” proudly describing its own A/B tests (e.g., removing profile photos from some profiles). The reaction was milder, because OkCupid admitted it itself. PR lesson: transparency always plays better than hiding.
Cambridge Analytica as a “late warning” — in 2018, when the CA case broke, many commentators pointed to “Emotional Contagion” as an early warning that had been ignored. If Facebook already knew in 2012 that it could manipulate the emotions of 700,000 people — what else?
PNAS never retracted the article, despite strong pressure. It is still indexed, cited (over 2,000 citations according to Google Scholar), and formally valid in the scientific literature. This shows the limits of academic self-regulation.
Adam Kramer after the scandal remained at Meta for years, but with a sharply reduced public profile. His Twitter account is private; he no longer publishes in PNAS.
Sheryl Sandberg in the book “Lean In” (2013), just a year before the scandal, wrote about the importance of authenticity and corporate ethics. After 2014, critics repeatedly reminded her of her words.
“Groupthink IRB”: documentation shows that the Cornell IRB review was conducted after the first questions from journalists — not before the experiment. A classic example of retroactive legitimation.
The irony of categorization: the article was placed in the “Social sciences” category of PNAS, not “Computer science”. Which suggests that PNAS itself understood it as a human subjects experiment, and not a data analysis.

Sources

Kramer, A.D.I., Guillory, J.E., Hancock, J.T., “Experimental evidence of massive-scale emotional contagion through social networks”, Proceedings of the National Academy of Sciences, vol. 111, no. 24, June 17, 2014, pp. 8788–8790. DOI: 10.1073/pnas.1320040111. URL: https://www.pnas.org/doi/10.1073/pnas.1320040111 (accessed: 2026-04-17)
Robinson Meyer, “Everything We Know About Facebook’s Secret Mood Manipulation Experiment”, The Atlantic, June 28, 2014. URL: https://www.theatlantic.com/technology/archive/2014/06/everything-we-know-about-facebooks-secret-mood-manipulation-experiment/373648/ (accessed: 2026-04-17)
Adam Kramer, public apology, Facebook, June 29, 2014. URL: https://www.facebook.com/akramer/posts/10152987150867796 (accessed: 2026-04-17)
Kashmir Hill, “Facebook Manipulated 689,003 Users’ Emotions For Science”, Forbes, June 28, 2014. URL: https://www.forbes.com/sites/kashmirhill/2014/06/28/facebook-manipulated-689003-users-emotions-for-science/ (accessed: 2026-04-17)
PNAS “Editorial Expression of Concern: Experimental evidence of massive-scale emotional contagion through social networks”, July 3, 2014. URL: https://www.pnas.org/doi/10.1073/pnas.1412469111 (accessed: 2026-04-17)
James Grimmelmann, “Illusion of Change” (blog), June 28, 2014 and subsequent posts. URL: https://laboratorium.net/ (accessed: 2026-04-17)
“When is academic research ethical research?”, Nature, vol. 510, June 26, 2014, p. 337. URL: https://www.nature.com/articles/510337a (accessed: 2026-04-17)
UK Information Commissioner’s Office, statement on the Facebook emotional contagion study, July 3, 2014 (press archive).
Federal Register, “Federal Policy for the Protection of Human Subjects — Final Rule”, January 19, 2017 (revised Common Rule). URL: https://www.federalregister.gov/documents/2017/01/19/2017-01058/federal-policy-for-the-protection-of-human-subjects (accessed: 2026-04-17)
Michelle Meyer, “Two Cheers for Corporate Experimentation: The A/B Illusion and the Virtues of Data-Driven Innovation”, Colorado Technology Law Journal, 2015.
Clay Johnson, analysis on Twitter, June 28–29, 2014 (archive).
Vindu Goel, “Facebook Tinkers With Users’ Emotions in News Feed Experiment, Stirring Outcry”, The New York Times, June 29, 2014. URL: https://www.nytimes.com/2014/06/30/technology/facebook-tinkers-with-users-emotions-in-news-feed-experiment-stirring-outcry.html (accessed: 2026-04-17)
Sheryl Sandberg, interview with NDTV in India (during her New Delhi visit), July 2, 2014. Cf. Washington Post report, July 2, 2014.
Washington Post, report of September 25, 2014 on the announcement by Douglas F. Gansler (Maryland AG, September 23–24, 2014) of an analysis of complaints by Prof. James Grimmelmann regarding the emotional contagion study.

Last updated: 2026-04-17 Card in database: A06_emotional_contagion.md