The Laboratorium
June 2014

This is an archive page. What you are looking at was posted sometime between 2000 and 2014. For more recent material, see the main blog at http://laboratorium.net

Parsing The Facebook Study’s Authorship and Review


I have been thinking a lot about the mechanics of how the Facebook emotional manipulation study was conducted, reviewed, and accepted for publication. I have found it helpful to gather in one place all of the various claims about who did what and what forms of review it received. I have bolded the relevant language.

What did the authors do?

PNAS authorship policy:

Authorship must be limited to those who have contributed substantially to the work. …

All collaborators share some degree of responsibility for any paper they coauthor. Some coauthors have responsibility for the entire paper as an accurate, verifiable report of the research. These include coauthors who are accountable for the integrity of the data reported in the paper, carry out the analysis, write the manuscript, present major findings at conferences, or provide scientific leadership to junior colleagues. Coauthors who make specific, limited contributions to a paper are responsible for their contributions but may have only limited responsibility for other results. While not all coauthors may be familiar with all aspects of the research presented in their paper, all collaborators should have in place an appropriate process for reviewing the accuracy of the reported results. Authors must indicate their specific contributions to the published work. This information will be published as a footnote to the paper. Examples of designations include:

  • Designed research
  • Performed research
  • Contributed new reagents or analytic tools
  • Analyzed data
  • Wrote the paper

An author may list more than one contribution, and more than one author may have contributed to the same aspect of the work.

From the paper:

Author contributions: A.D.I.K., J.E.G., and J.T.H. designed research; A.D.I.K. performed research; A.D.I.K. analyzed data; and A.D.I.K., J.E.G., and J.T.H. wrote the paper.

Cornell press release:

… According to a new study by social scientists at Cornell, the University of California, San Francisco (UCSF), and Facebook, emotions can spread among users of online social networks.

The researchers reduced the amount of either positive or negative stories that appeared in the news feed of 689,003 randomly selected Facebook users, and found that the so-called “emotional contagion” effect worked both ways.

“People who had positive content experimentally reduced on their Facebook news feed, for one week, used more negative words in their status updates,” reports Jeff Hancock, professor of communication at Cornell’s College of Agriculture and Life Sciences and co-director of its Social Media Lab. …

Cornell statement

Cornell University Professor of Communication and Information Science Jeffrey Hancock and Jamie Guillory, a Cornell doctoral student at the time (now at University of California San Francisco) analyzed results from previously conducted research by Facebook into emotional contagion among its users. Professor Hancock and Dr. Guillory did not participate in data collection and did not have access to user data. Their work was limited to initial discussions, analyzing the research results and working with colleagues from Facebook to prepare the peer-reviewed paper “Experimental Evidence of Massive-Scale Emotional Contagion through Social Networks,” published online June 2 in Proceedings of the National Academy of Science-Social Science.

Because the research was conducted independently by Facebook and Professor Hancock had access only to results – and not to any data at any time – Cornell University’s Institutional Review Board concluded that he was not directly engaged in human research and that no review by the Cornell Human Research Protection Program was required.

Adam Kramer’s statement for Facebook:

OK so. A lot of people have asked me about my and Jamie and Jeff’s recent study published in PNAS, and I wanted to give a brief public explanation. …

Regarding methodology, our research sought to investigate the above claim by very minimally deprioritizing a small percentage of content in News Feed (based on whether there was an emotional word in the post) for a group of people (about 0.04% of users, or 1 in 2500) for a short period (one week, in early 2012). … And we found the exact opposite to what was then the conventional wisdom: Seeing a certain kind of emotion (positive) encourages it rather than suppresses is.

What did the IRB do?

PNAS IRB review policy:

Research involving Human and Animal Participants and Clinical Trials must have been approved by the author’s institutional review board. … Authors must include in the Methods section a brief statement identifying the institutional and/or licensing committee approving the experiments. For experiments involving human participants, authors must also include a statement confirming that informed consent was obtained from all participants. All experiments must have been conducted according to the principles expressed in the Declaration of Helsinki.

Susan Fiske’s email to Matt Pearce:

I was concerned about this ethical issue as well, but the authors indicated that their university IRB had approved the study, on the grounds that Facebook filters user news feeds all the time, per the user agreement. Thus, it fits everyday experiences for users, even if they do not often consider the nature of Facebook’s systematic interventions. The Cornell IRB considered it a pre-existing dataset because Facebook continually creates these interventions, as allowed by the user agreement.

Having chaired an IRB for a decade and having written on human subjects research ethics, I judged that PNAS should not second-guess the relevant IRB.

I regret not insisting that the authors insert their IRB approval in the body of the paper, but we did check that they had it.

Fiske’s email to Adrienne LaFrance:

Their revision letter said they had Cornell IRB approval as a “pre-existing dataset” presumably from FB, who seems to have reviewed it as well in some unspecified way. (I know University regulations for human subjects, but not FB’s.) So maybe both are true.

Cornell’s statement (again):

Because the research was conducted independently by Facebook and Professor Hancock had access only to results – and not to any data at any time – Cornell University’s Institutional Review Board concluded that he was not directly engaged in human research and that no review by the Cornell Human Research Protection Program was required.

Kramer’s statement (again):

While we’ve always considered what research we do carefully, we (not just me, several other researchers at Facebook) have been working on improving our internal review practices. The experiment in question was run in early 2012, and we have come a long way since then. Those review practices will also incorporate what we’ve learned from the reaction to this paper.

The Facebook Emotional Manipulation Study: Sources


This post rolls up all of the major primary sources for the Facebook emotional manipulation study, along with selected news and commentary.


Paper:

  • “Experimental evidence of massive-scale emotional contagion through social networks” as PDF and as HTML (received Oct. 23, 2013, approved March 25, 2014, publication date June 17, 2014)

Authors:

Cornell:

UCSF: (Guillory became affiliated with UCSF only after the study was conducted)

*Human subjects policy

Facebook:

PNAS:

Common Rule:

Previous Facebook studies:

Journalism:

Commentary:

Misc.:

OK Cupid experiments:

Interviews with Christian Rudder about OKCupid experiments:

Ashley Madison study:

Facebook’s Modified Research Policy:

As Flies to Wanton Boys


Most recent update: 9:05 PM, Monday June 30

If you were feeling glum in January 2012, it might not have been you. Facebook ran an experiment on 689,003 users to see if it could manipulate their emotions. One experimental group had stories with positive words like “love” and “nice” filtered out of their News Feeds; another experimental group had stories with negative words like “hurt” and “nasty” filtered out. And indeed, people who saw fewer positive posts created fewer of their own. Facebook made them sad for a psych experiment.

I first saw the story on Facebook, where a friend picked it up from the A.V. Club, which got it from Animal, which got it from the New Scientist, which reported directly on the paper. It’s exploding across the Internet today (e.g. MetaFilter), and seems to be generating two kinds of reactions: outrage and shrugs. I tend more towards anger; let me explain why.

Facebook users didn’t give informed consent: The study says:

[The study] was consistent with Facebook’s Data Use Policy, to which all users agree prior to creating an account on Facebook, constituting informed consent for this research.

The standard of consent for terms of service is low. But that “consent” is a legal fiction, designed to facilitate online interactions. (See Nancy Kim and Margaret Jane Radin’s books for more.) It’s very different from informed consent, the ethical and legal standard for human subjects research (HSR). The Federal Policy for the Protection of Human Subjects, a/k/a the Common Rule, requires that informed consent include:

(1) A statement that the study involves research, an explanation of the purposes of the research and the expected duration of the subject’s participation, a description of the procedures to be followed, and identification of any procedures which are experimental;

(2) A description of any reasonably foreseeable risks or discomforts to the subject; …

(7) An explanation of whom to contact for answers to pertinent questions about the research and research subjects’ rights, and whom to contact in the event of a research-related injury to the subject;

(8) A statement that participation is voluntary, refusal to participate will involve no penalty or loss of benefits to which the subject is otherwise entitled, and the subject may discontinue participation at any time without penalty or loss of benefits to which the subject is otherwise entitled.

Facebook’s actual Data Use Policy contains none of these, only general statements that “we may use the information we receive about you … for internal operations, including troubleshooting, data analysis, testing, research and service improvement.” and “We give your information to the people and companies that help us provide, understand and improve the services we offer. For example, we may use outside vendors to … conduct and publish research.” Neither of these comes close to a “description of the procedures to be followed” or a “description of any reasonably foreseeable risks or discomforts,” and the Data Use Policy doesn’t even attempt to offer a contact for questions or an opt-out.

Federal law requires informed consent: To be sure, the Common Rule generally only applies to federally funded research, and Facebook is a private company. But that’s not the end of the story. The paper has three co-authors: Facebook’s Adam Kramer, but also Jamie Guillory from UCSF and Jeffrey Hancock from Cornell. UCSF and Cornell are major research universities and receive large sums of federal funding. Both of them have institutional review boards (IRBs), as required by the Common Rule: an IRB examines proposed research protocols to make sure they protect participants, obtain informed consent, and otherwise comply with ethical and legal guidelines.

I don’t know whether the study authors presented it to an IRB (the paper doesn’t say), but it strikes me as the sort of research that requires IRB approval. It further strikes me that the protocol as described is problematic, for the reasons described above. I don’t know whether I’m more afraid that the authors never obtained IRB approval or that an IRB signed off on a project that was designed to (and did!) make unsuspecting victims sadder.

The study harmed participants: The paper also argues:

[The study software] was adapted to run on the Hadoop Map/Reduce system (11) and in the News Feed filtering system, such that no text was seen by the researchers.

This claim misses the point. For an observational study, automated data processing is a meaningful way of avoiding privacy harms to research subjects. (Can robot readers cause a privacy harm? Bruce Boyden would say no; Samir Chopa would say yes.) But that is because in an observational study, the principal risks to participants come from being observed by the wrong eyes.

This, however, was not an observational study. It was an experimental study—indeed, a randomized controlled trial—in which participants were treated differently. We wouldn’t tell patients in a drug trial that the study was harmless because only a computer would ever know whether they received the placebo. The unwitting participants in the Facebook study were told (seemingly by their friends) for a week either that the world was a dark and cheerless place or that it was a saccharine paradise. That’s psychological manipulation, even when it’s carried out automatically.

This is bad, even for Facebook: Of course, it’s well know that Facebook, like other services, extensively manipulates what it shows users. (For recent discussions, see Zeynep Tufekci, Jonathan Zittrain, and Christian Sandvig). Advertisers and politicians have been in the emotional manipulation game for a long time. Why, then, should this study—carried out for nobler, scientific purposes—trigger a harsher response?

One reason is simply that some walks of life are regulated, and Facebook shouldn’t receive a free pass when it trespasses into them simply because it does the same things elsewhere. Facebook Beacon, which told your Facebook friends what you were doing on other sites, was bad everywhere but truly ugly when it collided with the Video Privacy Protection Act. So here. Whatever you think of Facebook’s ordinary marketing-driven A/B testing is one thing: what you think of it when it hops the fence into Common Rule-regulated HSR is quite another. Facebook has chosen to go walking in a legal and ethical minefield; we should feel little sympathy when it occasionally blows up. (That said, insisting on this line would simply drive future research out of the academy and into industry, where our oversight over it will be even weaker. Thus …)

A stronger reason is that even when Facebook manipulates our News Feeds to sell us things, it is supposed—legally and ethically—to meet certain minimal standards. Anything on Facebook that is actually an ad is labelled as such (even if not always clearly.) This study failed even that test, and for a particularly unappealing research goal: We wanted to see if we could make you feel bad without you noticing. We succeeded. The combination of personalization and non-rational manipulation may demand heightened legal responses. (See, e.g., Ryan Calo, or my thoughts on search engines as advisors.)

The real scandal, then, is what’s considered “ethical.” The argument that Facebook already advertises, personalizes, and manipulates is at heart a claim that our moral expectations for Facebook are already so debased that they can sink no lower. I beg to differ. This study is a scandal because it brought Facebook’s troubling practices into a realm—academia—where we still have standards of treating people with dignity and serving the common good. The sunlight of academic practices throws into sharper relief Facebook’s utter unconcern for its users and for society. The study itself is not the problem; the problem is our astonishingly low standards for Facebook and other digital manipulators.

This is a big deal: In 2006, AOL released a collection of twenty million search queries to researchers. Like the Facebook study authors, AOL thought it was protecting its users: it anonymized the users’ names. But that wasn’t sufficient: queries like “‘homes sold in shadow lake subdivision gwinnett county georgia” led a reporter straight to user No. 4417749. Like Facebook, AOL had simply not thought through the legal and ethical issues involved in putting its business data to research purposes.

The AOL search-query release became known as the “Data Valdez” because it was a vivid and instantly recognizable symbol of the dangers of poor data security. It shocked the public (and the industry) into attention, and put search privacy on the map. I predict, or at least I hope, that the Facebook emotional manipulation study will do the same for invisible personalization. It shows, in one easy-to-grasp lesson, both the power Facebook and its fellow filters hold to shape our online lives, and the casual disdain for us with which they go about it.

UPDATE: The study was presented to an IRB, which approved it “on the grounds that Facebook filters user news feeds all the time, per the agreement.” See @ZLeeily, with hat tips to Kashmir Hill and @jon_penney.

UPDATE: Another @jon_penney pickup: it appears that the study itself was federally funded. Cornell amended the press release to say that the claim of federal funding was in error.

UPDATE: Kashmir Hill reports:

Professor Susan Fiske, the editor at the Proceedings of the National Academy of Sciences for the study’s publication, says the data analysis was approved by a Cornell Institutional Review Board but not the data collection. “Their revision letter said they had Cornell IRB approval as a ‘pre-existing dataset’ presumably from Facebook, who seems to have reviewed it as well in some unspecified way,” writes Fiske by email.

UPDATE: For much more on the IRB legal issues, see this detailed post by Michelle Meyer. She observes that the Common Rule allows for the “waiver or alteration” of informed consent for research that poses “minimal risk” to participants. The crucial issue there is whether the study “could not practicably be carried out without the waiver or alteration.” Meyer also has an extended discussion of whether the Common Rule apples to this research—following the Cornell restatement, it is much less clear that it does.

UPDATE: I’ve created a page of primary sources related to the study and will update it as more information comes in.

Google Books Round 86: Libraries Win Yet Again


The Second Circuit’s decision in Authors Guild v. HathiTrust is out. This, as a reminder, is the offshoot of the Google Books litigation in which the Authors Guild inexplicably sued Google’s library partners. The trial judge, Harold Baer, held for the libraries in 2012 in a positively exuberant opinion:

I cannot imagine a definition of fair use that would not encompass the transformative uses made by Defendants’ MDP [Mass Digitization Project] and would require that I terminate this invaluable contribution to the progress of science and cultivation of the arts that at the same time effectuates the ideals espoused by the ADA.

The Second Circuit’s opinion drops the grand rhetoric, but otherwise the bottom line is basically the same: mass digitization to make a search engine is fair use, and so is giving digital copies to the print-disabled. The opinion on appeal is sober, conservative, and to the point; it is the work of a court that does not think this is a hard case.

On full-text search:

  • Factor 1: “[T]he creation of a full‐text searchable database is a quintessentially transformative use” because it serves a “new and different function.” Authors write to be read, not to be searched.
  • Factor 2: The nature of the copyrighted work fades into irrelevance for transformative uses.
  • Factor 3: Since full-text search requires copying full books, the copying isn’t excessive in light of the use. True, HathiTrust makes four copies of each book, two live and two in tape backup, but those are appropriate precautions against Internet outages and natural disasters. (It’s nice to see a court recognize that strict copy-counting is a fool’s errand in light of modern IT; better to focus, as the court here does, on the uses those copies enable.)
  • Factor 4: “[T]he full‐text‐search use poses no harm to any existing or potential traditional market … .” Book reviews do not substitute for sales of a book, even when they convince readers not to buy the book; so here. There is no lost licensing market because full-text search is not a substitute for books in the first place. (No citation to American Geophysical!) And while the Authors Guild says there’s a risk of a security breach, saying so doesn’t make it so: the harm from a hypothetical breach is pure speculation.

On print-disabled access:

  • Factor 1: Providing access to the print-disabled is not transformative: “By making copyrighted works available in formats accessible to the disabled, the HDL [HathiTrust Digital Library] enables a larger audience to read those works, but the underlying purpose of the HDL’s use is the same as the author’s original purpose.” But providing such access is still a favored use: there is a national policy of promoting access, reflected in the Chafee Amendment and recognized by the Supreme Court.
  • Factor 2: Irrelevant again, even though the use isn’t transformative. (Factor 2 never matters for published expressive works.)
  • Factor 3: The scanned images—and not just the OCR’ed text—are useful to print-disabled readers. Some readers are print-disabled because they need greater magnification or stronger color contrast than paper provides, others because they can’t turn pages. Scanned images help them both. (It’s nice to see a court take the diversity of disabilities seriously; Dan Goldstein’s advocacy here clearly helped.)
  • Factor 4: There is no market for selling books to the print-disabled; only a small percentage of books are published in accessible formats and even for those authors typically forego their royalties. (The Authors Guild’s war against text to speech has come back to bite it.)

These holdings merely affirm the District Court’s conclusions, but they are still a big deal. The Second Circuit’s decisions are binding precedent in New York, the nation’s publishing capital, and are highly influential beyond. Five judges have now upheld the legality of scanning books to make a search engine; none has disagreed.

The other major points in the opinion all consist of declining to decide:

  • The Authors Guild lacks standing to sue on behalf of its members. The case continues, thanks to the international organizations and the individual plaintiffs, but ouch. By pressing the Google Books cases, the Authors Guild has undercut its ability to take legal action on behalf of “authors” in general. In a real sense, it is legally weaker than when the case started.
  • Preservation uses aren’t ripe for consideration because the court has already held that hanging on to four copies is fully justified by the operation demands of providing full-text search. That only leaves printing replacement copies for lost or damaged ones when they’re unavailable for purchase at a fair price, but since it’s not clear whether or when that would happen—let alone whether it would happen to one of the remaining plaintiffs’ books—the issue isn’t ripe to decide.
  • Since Michigan has suspended the orphan works project (showing orphaned works to non-disabled patrons) and has no plans to reinstate it in the same form, those issues aren’t ripe either. The libraries dodged a bullet here; if they want to try again, it will be on terms of their choosing.

The opinion is a green light for library search engine digitization. It is an even greener light for making books and other works accessible to the disabled. And there was great rejoicing at the DPLA and the Internet Archive. There is not very much new in the opinion, but its very lack of novelty sends a strong signal that these uses are now clearly established.

What next? The Authors Guild could ask for rehearing, or petition for certiorari. I personally don’t like those odds, but I have never really understood the Guild’s decision-making process around this case, so who knows? The opinion sends a strong signal that the case against Google, also on appeal to the Second Circuit, is also likely to go in favor of scanning. At the very least, if the two cases are to be distinguished, it will have to be on narrow grounds: that Google makes commercial uses or shows snippets. Even that would provide clear guidance for digitizers. The holding may also cast a shadow on other search, education, and access cases, for example the Georgia State e-reserves case.

Solar Roadblocks


There are many reasons to love George Takei. But this is not one of them:

Takei’s tweet helped the Solar Roadways project’s Indegogo fundraising campaign blow past its million-dollar goal. Their plan is to replace asphalt road surfaces with durable solar panels with embedded LEDs and sensors, turning highways into smart power-generating grids that can melt snow and give drivers safety warnings. Each individual hexagonal panel is capable of cranking out only a few watts, but if you do out the math on replacing all our highways with the the panels, it comes out to a ludicrous sum, well more than the United States’s entire current energy consumption.

It pains me, then, to say that the idea behind Solar Roadways isn’t just crazy; it’s obviously crazy. It’s Troy Hurtubise crazy; it’s Dr. Evil crazy. All the hype around solar panels and LEDs simply disguises the fact that Solar Roadways fundamentally misunderstands what a road is.

A road is a system for distributing moving loads into the ground.

That is its one indispensable job: to allow people and vehicles to travel atop it while absorbing the forces they create. I don’t claim to be a civil engineer. But I know enough high-school physics to be capable of asking questions whose answers are nowhere to be found on the Solar Roadways site, in any technical documentation, or anywhere at all on the Internet (so far as I can tell). Solar Roadways has simply not attempted to address the bread-and-butter engineering problems that highway builders have spent decades dealing with. Here are a few:

  • What do the solar panels rest on? The sublayer beneath them has to be made of something. That something will receive the forces transmitted downwards through the panels, and that something will degrade over time. To fix it, you’ll have to remove and restore the panels.
  • Cars and trucks will put their weight unevenly, on different parts of the panels. How resistant are the panels to bending? To shearing? These are different from the simpler question that Solar Roadways does discuss: how much weight they can take before being crushed.
  • Cars and trucks will accelerate and brake and push against the air; they have to push off against something. Solar Roadways has thought about traction. But receiving the force from vehicles is only half the problem, because they must also transmit the force to the sublayer. How are they anchored to it? How will the anchors deal with the immense lateral force of a braking tractor-trailer?
  • Water will get between and below the panels, especially in climates where the panels are supposed to melt snow. What happens when that water expands as it freezes? What happens when it freezes and unfreezes repeatedly?

I am not suggesting that these are insurmountable engineering challenges. We live in an age of near miracles. I am simply suggesting that they are challenges, and that they are obviously challenges. Not to see them fleshed out in the slightest is deeply discouraging, because it means that Solar Roadways is not approaching a gigantic engineering problem as an engineering problem. Even the Hyperloop—the Hyperloop!—came with a design document that tried (if not always successfully) to think through the engineering issues. When your futuristic transport technology is bigger vaporware than the Hyperloop, you have a problem.

It is not as though these are exotic problems, like building a quantum computer. These are familiar problems; they are the bread and butter of highway engineers. But no one has asked in detail, “What problems do road-builders currently solve, and how will solar roadways deal with these same issues?” Asphalt, for all its other issues, distributes moving loads quite well for its price point. To work as roads, Solar Roadways will need to replicate that success. Leave the solar panels and electronics aside for a moment: if building road surfaces out of thick glass was a good idea in its own right, we would be much more familiar with glass roads.

But assume even that these issues are all in the end solvable. Will they be solvable at cost? Almost certainly not. Solar power engineering faces its own significant design constraints. We are making progress in bringing the cost down, but still this is hard stuff. These are completely different design constraints than highway engineering faces. Why on earth would you insist on solving both sets of problems simultaneously in the same surface, if you didn’t have to?

There is no plausible future in which solar-panel roads make more sense than solar panels plus roads. There are plenty of other places to put solar panels, and plenty of other ways to make highways smarter. Solar sidewalks. Solar medianstrips. Maybe these are also terrible ideas. But they are unambiguously better ideas than solar roadways.

Science and engineering, done right, can be beautiful and amazing. But it doesn’t run in the other direction. Something can be as cool and as awesome as solar frickin roadways, and still not work as science. I understand the impulse that made people open their wallets to support this appeal to progressive technology as a solution to humanity’s self-inflicted woes. But science is the pursuit of truth, not truthiness, and solar roadways are scienciness, not science.