Guest Post: A Facebook Apologia

July 8, 2014

As you might have heard, Face­book recently released a study in col­lab­o­ra­tion with researchers at Cor­nell Uni­ver­sity on the spread of emo­tional sen­ti­ment through the social net­work. It has spurred a huge, media-​​fueled debate among the main­stream public. Below is a post that Brian Keegan, a post-​​doctoral researcher in North­eastern pro­fessor David Lazer’s lab, wrote on his per­sonal blog explaining the sci­ence, the debate, and its poten­tial impacts on the field.

Last week, the Pro­ceed­ings of the National Academy of Sci­ence (PNAS) pub­lished a study that con­ducted a large-​​scale exper­i­ment on Face­book. The authors of the study included an industry researcher from Face­book as well as aca­d­e­mics at the Uni­ver­sity of Cal­i­fornia, San Fran­cisco and Cor­nell Uni­ver­sity. The study employed an exper­i­mental design that reduced the amount of pos­i­tive or neg­a­tive emo­tional con­tent in 689,000 Face­book users’ news feeds to test whether emo­tions are contagious. The study has since spawned a sub­stan­tial con­tro­versy about the methods used, extent of its reg­u­la­tion by aca­d­emic insti­tu­tions’ review board, the nature of par­tic­i­pants’ informed con­sent, the ethics of the research design itself, and the need for more explicit opt-​​in pro­ce­dures.

In the face of even-​​tempered thinking from a gath­ering mob, I want to defend the exe­cu­tion and impli­ca­tions of this study. Others have also made sim­ilar argu­ments [1,2,3], I guess I’m just a slow blogger. At the outset, I want to declare that I have no direct stake in the out­come of this brouhaha. How­ever, I do have pro­fes­sional and per­sonal rela­tion­ships with sev­eral mem­bers of the Face­book Data Sci­ence team (none of whom are authors on the study), although the entirety of this post reflects only public infor­ma­tion and my opin­ions alone.

First, as is common in the ini­tial reporting sur­rounding on sci­en­tific find­ings, there was some mis­in­for­ma­tion around the study that was greatly mag­ni­fied. These early crit­i­cisms claimed the authors mis-​​represented the size of the observed effects (they didn’t) or the research wasn’t reviewed by the aca­d­emic boards charged with human sub­jects pro­tec­tion (it was). There is like­wise a per­ni­cious ten­dency for the sci­en­tific con­cept of exper­i­mental manip­u­la­tion to be mis­in­ter­preted as the homo­phone implying decep­tion and chi­canery: there is no inherent mali­cious­ness in ran­domly assigning par­tic­i­pants to con­di­tions for exper­i­mental study. Other reporting on the story has sen­sa­tion­al­is­ti­cally implied users were sub­jected to injec­tions of neg­a­tive emo­tional con­tent so their resulting depres­sion could be more fully quan­ti­fied. In reality, the study actu­ally only with­held either pos­i­tive or neg­a­tive con­tent from users, which resulted in users seeing more of posts they would have seen anyway. In all of these, the hys­teria sur­rounding a “Face­book manip­u­lates your emo­tions” or is “trans­mit­ting anger” story got well ahead of any sober reading of the research reported by the authors in the paper.

Second on the sub­stance of the research, there are still serious ques­tions about the validity of method­olog­ical tools used , the inter­pre­ta­tion of results, and use of inap­pro­priate con­structs. Pres­ti­gious and com­pet­i­tive peer-​​reviewed jour­nals like PNAS are not immune from pub­lishing studies with half-​​baked analyses. Pre-​​publication peer review (as this study went through) is impor­tant for serving as a check against faulty or improper claims, but post-​​publication peer review of scrutiny from the sci­en­tific community—and ide­ally replication—is an essen­tial part of sci­en­tific research. Pub­lishing in PNAS implies the authors were seeking both a wider audi­ence and a height­ened level of scrutiny than pub­lishing this paper in a less promi­nent outlet. To be clear: this study is not without its flaws, but these debates, in of them­selves, should not be taken as evi­dence that the study is irrec­on­cil­ably flawed. If the bar for pub­li­ca­tion is antic­i­pating every poten­tial objec­tion or addressing every method­olog­ical lim­i­ta­tion, there would be pre­cious little schol­ar­ship for us to dis­cuss. Debates about the con­structs, methods, results, and inter­pre­ta­tions of a study are cru­cial for syn­the­sizing research across dis­ci­plines and increasing the quality of sub­se­quent research.

Third, I want to move to the issue of epis­te­mology and framing. There is a pro­found dis­con­nect in how we talk about the ways of knowing how sys­tems like Face­book work and the ways of knowing how people behave. As users, we expect these sys­tems to be respon­sive, effi­cient, and useful and so com­pa­nies employ thou­sands of engi­neers, product man­agers, and usability experts to create seam­less expe­ri­ences. These user expe­ri­ences require diverse and iter­a­tive methods, which include A/​B testing to com­pare users’ pref­er­ences for one design over another based on how they behave. These tests are per­va­sive, active, and on-​​going across every con­ceiv­able online and offline envi­ron­ment from couponing to product recommendations. Creating expe­ri­ences that are “pleasing”, “intu­itive”, “exciting”, “over­whelming”, or “sur­prising” reflects the fun­da­men­tally psy­cho­log­ical nature of this work: every A/​B test is a psych experiment.

Some­where deep in the fine print of every loy­alty card’s terms of ser­vice or online account’s pri­vacy policy is some lan­guage in which you con­sent to having this data used for “trou­bleshooting, data analysis, testing, research,” which is to say, you and your data can be sub­ject to sci­en­tific obser­va­tion and exper­i­men­ta­tion. Whether this con­sent is “informed” by the par­tic­i­pant having a con­scious under­standing of impli­ca­tions and con­se­quences is a very dif­ferent ques­tion that I sus­pect few com­pa­nies are pre­pared to defend. But why does a framing of “sci­en­tific research” seem so much more prob­lem­atic than con­tributing to “user expe­ri­ence”? How is pub­lishing the results of one A/​B test worse than knowing nothing of the thou­sands of invisble tests? They reflect the same sub­stan­tive ways of knowing “what works” through the same well-​​worn sci­en­tific methods.

Fourth, there has been no sub­stan­tive dis­cus­sion of what the design of informed con­sent should look like in this con­text. Is it a blanket opt-​​in/​out to all exper­i­men­ta­tion? Is con­sent needed for every single A/​B iter­a­tion or only those intended for sci­en­tific research? Is this choice buried along­side all the other com­plex pri­vacy but­tons or are users expected to manage pop-​​ups requesting your participation? I sus­pect the omnipresent secu­rity dia­logues that Win­dows and OS X have adopted to warn us against installing soft­ware have done little to reduce risky behavior. Does adding another layer of com­plexity around informed con­sent improve the cur­rent anx­i­eties around man­aging com­plex pri­vacy settings? How would users go about dif­fer­en­ti­ating offi­cial requests for informed con­sent from abu­sive apps, spam­mers, and spoofers? Who should be charged with enforcing these rules and who are they in turn account­able to? There’s been pre­cious little on designing more informed con­sent archi­tec­tures that bal­ance usability, plat­form affor­dances, and the needs of researchers.

Fur­ther­more, we might also con­sider the ethics of this nascent socio-​​technical NIM­BYism. Researchers at Penn State have looked at the design of pri­vacy autho­riza­tion dia­logues for social net­works but found that more fine-​​grained con­trol over dis­clo­sure reduced adop­tion levels. We demand ever more respon­sive and pow­erful sys­tems while cir­cum­scribing our con­tri­bu­tions but demanding ben­e­fits from other’s con­tri­bu­tions. I image the life of such sys­tems would be poor, nasty, brutish, and short. Do more obtru­sive inter­ven­tions or incom­plete data col­lec­tion in the name con­ser­v­a­tive inter­pre­ta­tions of informed con­sent pro­mote better sci­ence and other public goods? What are the spe­cific harms that we should strive to limit in these sys­tems and how might we re-​​tailor 40 year old poli­cies to these ends?

I want to wrap up by shifting the focus of this con­ver­sa­tion from debates about a study that was already done to what should be done going for­ward. Some of the more extreme calls I’ve seen have advo­cated for aca­d­emic soci­eties or insti­tu­tions to inves­ti­gate and dis­ci­pline the authors, others have called for embar­going studies using Face­book data from schol­arly pub­li­ca­tion, and still others have encour­aged Face­book employees to quit in protest of a single study. All this man­ning of bar­ri­cades strikes me as a grave over-​​reaction that could have calami­tously chilling effects on sev­eral dimen­sions. If our over­riding social goal is to min­i­mize real or poten­tial harm to par­tic­i­pants, what best accom­plishes this going forward?

Cer­tainly expelling Face­book from the “com­mu­nity of scholars” might damage its ability to recruit researchers. But are Face­book users really made safer by replacing its cur­rent crop of data sci­en­tists who have superla­tive social sci­ence cre­den­tials with engi­neers, mar­keters, and product man­agers trying to ride method­olog­ical bulls they don’t under­stand? Does Face­book have greater out­side insti­tu­tional account­ability by closing down aca­d­emic col­lab­o­ra­tions and shut­ting papers out from peer review and pub­li­ca­tion? Are we better able to know the poten­tial influ­ence Face­book wields over our emo­tions, rela­tion­ships, and pol­i­tics by dis­cour­aging them from pub­licly dis­closing the tools they have developed? Is raising online mobs to attack industry researchers con­ducive to starting dia­logues to improve their processes for informed con­sent? Is pub­licly under­mining other sci­en­tists the right strategy for pro­moting evidence-​​based policy-​​making in an increas­ingly hos­tile polit­ical climate?

Need­less to say, this episode speaks for the need for rap­proche­ment and sus­tained engage­ment between industry and aca­d­emic researchers. If you care about research ethics, informed con­sent, and well-​​designed research, you want com­pa­nies like Face­book deeply embedded within and respon­sible to the broader research com­mu­nity. You want the values of social sci­en­tists to influ­ence the prac­tice of data sci­ence, engi­neering, user expe­ri­ence, and mar­keting teams. You want the campus to be open to vis­iting aca­d­emic researchers to explore, col­lab­o­rate, and repli­cate. You want industry research to be held to academia’s more strin­gent stan­dards of human sub­jects pro­tec­tion and reg­u­larly shared through peer-​​reviewed publication.

The Face­book emo­tional con­ta­gion study demands a re-​​evaluation of pre­vailing research ethics, design values, and algo­rithmic powers in mas­sive net­worked architectures. But the cur­rent reac­tion to this study can only have a chilling effect on this debate by removing a unique form respon­sible dis­clo­sure through aca­d­emic col­lab­o­ra­tion and pub­lishing. This study is guar­an­teed to serve as an impor­tant case study in the pro­fes­sion­al­iza­tion of data sci­ence. But aca­d­emic researchers should make sure their reac­tions do not unin­ten­tion­ally inoc­u­late industry against the values and per­spec­tives of social inquiry.