E-mail:
Jack Balkin: jackbalkin at yahoo.com
Bruce Ackerman bruce.ackerman at yale.edu
Ian Ayres ian.ayres at yale.edu
Corey Brettschneider corey_brettschneider at brown.edu
Mary Dudziak mary.l.dudziak at emory.edu
Joey Fishkin joey.fishkin at gmail.com
Heather Gerken heather.gerken at yale.edu
Abbe Gluck abbe.gluck at yale.edu
Mark Graber mgraber at law.umaryland.edu
Stephen Griffin sgriffin at tulane.edu
Jonathan Hafetz jonathan.hafetz at shu.edu
Jeremy Kessler jkessler at law.columbia.edu
Andrew Koppelman akoppelman at law.northwestern.edu
Marty Lederman msl46 at law.georgetown.edu
Sanford Levinson slevinson at law.utexas.edu
David Luban david.luban at gmail.com
Gerard Magliocca gmaglioc at iupui.edu
Jason Mazzone mazzonej at illinois.edu
Linda McClain lmcclain at bu.edu
John Mikhail mikhail at law.georgetown.edu
Frank Pasquale pasquale.frank at gmail.com
Nate Persily npersily at gmail.com
Michael Stokes Paulsen michaelstokespaulsen at gmail.com
Deborah Pearlstein dpearlst at yu.edu
Rick Pildes rick.pildes at nyu.edu
David Pozen dpozen at law.columbia.edu
Richard Primus raprimus at umich.edu
K. Sabeel Rahmansabeel.rahman at brooklaw.edu
Alice Ristroph alice.ristroph at shu.edu
Neil Siegel siegel at law.duke.edu
David Super david.super at law.georgetown.edu
Brian Tamanaha btamanaha at wulaw.wustl.edu
Nelson Tebbe nelson.tebbe at brooklaw.edu
Mark Tushnet mtushnet at law.harvard.edu
Adam Winkler winkler at ucla.edu
Algorithms, and recommendation algorithms in
particular, are deeply ingrained in our networked public sphere. Facebook recommends us new friends, pages to
like or groups to join, Google websites that fit our search interests, Twitter
people to follow or topics to check out, Amazon products to buy, Spotify music
to listen to, and YouTube videos to watch.
According to Alexa, the most visited websites in
the U.S. are Google, YouTube, Amazon, Yahoo, and Facebook. Recommendation
algorithms are an integral part of each of those websites. As Eli Pariser, Safiya Umoja Noble, Frank Pasquale, and many others have argued, there are myriad
reasons to be concerned about algorithms’ integral role in our dailylives. These include the creation of
homogenous communities without our knowledge; reproduction of racism and
misogyny; and the concealment of algorithmic decisions to begin with. Against
this background, as well as my own research on the far-right, disinformation,
and algorithms, I argue that social media companies need to deactivate their
recommendation algorithms in the political sphere. I structure this demand
around our knowledge of how machine learning works, the pitfalls of automated
content curation, how corporate goals can run counter to the public good, as
well as findings from my own research.
When thinking about recommendation algorithms,
we need to think about machine learning, and statistical probability models.
After all, this is what algorithms are. Yet, as British statistician George Box famously said: “All
models are wrong, but some are useful.” As Momin Malik
highlights convincingly, these models are approximations, and there are
numerous ways in which they can fail or have shortcomings. On a more applied
note, Harini Suresh
and John Guttag identified five biases that compromise
algorithms: historical bias, representation bias, measurement bias, aggregation
bias, and evaluation bias. These biases can lead to problematic outcomes that
might reproduce issues such as racism, misogyny, and more. In short:
algorithms, no matter how good, will always have limitations.
This
problem is only exacerbated when focused on the supply side — i.e., the content
on a platform. As my colleague Adrian Rauchfleisch and I have
highlighted, far-right content that heavily features racism
or disinformation plays a significant role in the German as well as the
American YouTube sphere. When comparing the prominence of far-right actors on
YouTube with general findings for the U.S. networked public sphere, the
extreme political fringe seems overrepresented on YouTube. Similarly, in
Germany, the YouTube political community mostly consists of
far-right and conspiratorial channels. This, then, highlights that political
communities on YouTube are hardly representative of the general media discourse
and seem to favor more radical voices.
But as Michael Golebiewski and Danah Boyd
highlight, even if that were not the case, algorithms face an inherent issue:
data voids. Data voids are, in short, gaps in the content that a platform can
recommend. For example, this can occur when a specific search term suddenly
gains popularity. These data voids, however, can be abused by malicious actors
who want to spread disinformation; recommendation algorithms cannot not recommend content. Indeed, they are
limited to content on their platforms. This inherent need to recommend thus can
feature harmful content; this is especially so when content on the platform is
already harmful. As my co-authors and I show in a forthcoming study on Zika in
Brazil, even when YouTube curated the search results for videos on the Zika
virus, misinformation was still present throughout the results and
recommendations. This, then, indicates that even when platforms attempt to
curate their recommendations, algorithms will nevertheless uncover and
recommend harmful content.
Add to
this that algorithms that are usually the property of companies and thus, as
Pasquale, highlights “black boxes.” This means that we can inherently only see,
measure, or interact with an algorithm’s output and can only guess on how the
algorithm ended up with its final recommendations. While every now and then we
get an idea of some of the factors that contribute to a platform’s
algorithms, the general audience, as well as the platform’s content creators, are
left in the dark. On some platforms like YouTube (but also recently TikTok),
the algorithm has thus even a “celestial” quality, as content creators’ success is dependent on it.
I argue,
however, that we don’t need to know what goes into the algorithm to understand
that their objectives are at odds with the public good and a utopian version of the public sphere.
Indeed, from everything that we know, algorithms are optimized on user
behavior and especially on how much time is spent on the platform. And while
companies profit off of users’ prolonged stays on their platforms, it is
unlikely that users are profiting to the same extent.
Indeed,
what keeps people engaged, i.e., viewing, commenting, etc., can to some extent
be traced back to negative content. As we know from studies on user comments, people
tend to write user comments that are more uncertain, negative, and
controversial. In addition, there is a reason why YouTube’s study focuses
on users’ satisfaction; presumably because engagement on YouTube is not driven
by content that is agreeable but rather controversial. In other words: What is
good for the company is not necessarily in service of the public good.
Which
brings me to my final point: the effect of algorithms. Indeed, little is known
so far in terms of the effect of recommendation algorithms. Yet, a study that
I conducted looked at the user comments in over 100 German
far-right channels and examined whether we could identify activity patterns
over time. Indeed, we were able to show that the community grew more central
over time, indicating that the users that, at first, only commented under one
channel eventually also commented under videos from other related channels. And
while we don’t know whether this finding can be explained with YouTube’s
algorithms alone, it is important to note that YouTube claims its algorithms
drive 70% of the traffic on the site. In this context, we have argued that YouTube’s recommendation algorithms can
cause a digital Thomas theorem that normalizes radical content and can,
effectively, nudge people towards more problematic and disinforming content.
Recommendation algorithms are, in my opinion,
unfixable. There is no doubt that they, in general, work and can be quite
useful. Indeed, most content on Facebook or YouTube is not political and, in
these contexts, one might have fewer issues with recommendation algorithms.
Yet, even in these supposedly benign contexts, algorithms can cause harm. While conducting research on YouTube in Brazil, my
co-authors and I stumbled on what The New
York Times eventually called “On YouTube’s Digital Playground, an Open Gate for Pedophiles.”
While recommendation algorithms are supposedly
neutral, neither the people creating them nor the people using or training them
are. For YouTube’s recommendation algorithm, people watching content is people
watching content and the algorithm attempts to optimize on their viewing
behavior; in this case, videos of children. And while this might be an extreme
example of a recommendation algorithm causing real harm, YouTube acted swiftly and deactivated its recommendation
algorithm when videos included children.
In this piece I am arguing for a similar step
for political content. As I have shown above, no matter how much work platforms
pour into their algorithms, they will always have limitations and will always
need curation. Combining these imperfect algorithms with a highly skewed group
of content creators, then, must end badly. No matter how much you tweak and
optimize the algorithms, if content is problematic, the algorithms will recommend
it. Deliberate attempts at “gaming” the algorithm and pushing disinformation,
coupled with humans being drawn to controversial and negative content, means
that as long as recommendation algorithms exist, problematic content will
surface and be recommended. This is especially so if algorithms are designed to
keep people on the platform, even when it is not for their own good. Finally,
these algorithms can contribute to the normalization of extreme and
disinforming content and nudge people to more radical communities.
I am not saying that these platforms should only
remove specific recommendations. No. I argue that we need to get rid of all
recommendations for political content. This is not about whether content can or
cannot exist on a platform. This is about whether recommendation algorithms can
be “saved.” I argue that they can’t and only demand platforms to take the same
steps that YouTube did when made aware of their algorithms pushing videos of
children to pedophiles: Deactivate the recommendation algorithms.
Jonas Kaiser is an assistant professor of communication, journalism, and media at Suffolk University.