In advance of his talk on online deliberation this afternoon, former Ash Center Democracy Fellow Alfred Moore shares some preliminary results from a project on online commenting that he conducted with his colleagues John Naughton and Rolf Fredheim. Using data from Huffington Post comment sections over time, Moore and his colleagues uncover interesting trends that debunk the assumption that civility necessarily increases with the decline of anonymity in online commenting spaces.
By Alfred Moore
A few years ago Facebook’s then marketing director, Randi Zuckerberg, said during a panel discussion on social media that ‘anonymity on the internet has to go away.’ ‘People behave a lot better when they have their real names down,’ she continued. ‘I think people hide behind anonymity and they feel like they can say whatever they want behind closed doors.’
Google’s Eric Schmidt, speaking in Edinburgh a month later, agreed: ‘The internet would be better if we had an accurate notion that you were a real person as opposed to a dog, or a fake person, or a spammer.’
Drawing on the same intuition, the Huffington Post a couple of years later stopped accepting anonymous comments, to reduce trolling and verbal abuse.
This all sounds reasonable enough. The hope is that real-name environments would limit abusive behaviour by binding users to discursive norms rooted in community. Would you talk like that in front of your mother?
The problem is that real-name environments, in which people tend to know each other and share a broad set of perspectives, values and opinions, create their own problems of conformity and social pressure.
So we circle around two contradictory normative positions. Anonymity is valuable because it enables expression free from fear of repercussions. But anonymity is destructive because it enables expression free from fear of repercussions. The same feature that enables a teenager from a repressive religious community to talk freely about his sexuality without fear of exposure also enables cruel and abusive responses that may inhibit such expressions.
As Zuckerberg’s comment illustrates, we seem to face a simple trade-off between the benefits and drawbacks of real name and anonymous environments. If you want users to be able to express themselves without fear of retribution or pressures towards conformity, you have to accept that they may use that freedom to be cruel and abusive.
And if you want to bind users to the norms of community, you have to accept the risk that users will censor themselves, seek to avoid conflict, and be subject to the pressures and expectations associated with their offline social identities.
We think anonymity is more complex, and that this trade-off between anonymity and real-name environments is not quite what it seems. We distinguish durability, or the ease or difficulty with which identities can be acquired or changed, from connectedness, which involves bridging and linking communication across different social contexts.
Our claim is that what’s bad about anonymity is weak durability. Where people can change or create new identities, communicative accountability is easily evaded. What’s bad about real name environments, on the other hand, is strong connectedness, which can reproduce offline power dynamics.
This distinction gives us three options instead of two: pure anonymity (not durable, not connected); pseudonymity (durable, not connected); and real-name (durable, connected).
These types are not perfectly recreated in any particular online environment, but the changes to the commenting architecture made by the Huffington Post over two years from January 2013 enabled us to take a look at something pretty close.
We gathered 50m comments on news articles featured on the Huffington Post’s front page from January 2013 to January 2015, from which we selected a smaller sample for closer analysis.
Over this time period, the Huffington Post changed its commenting architecture, creating three distinct phases. The first of these phases allowed anonymous commenting. At this time, the platform experienced aggressive ‘trolling’ and the use of multiple accounts.
In December 2013 HuffPo moved to regulate its forum by requiring users to authenticate their accounts through Facebook. This created a division: HuffPo could be fairly sure that each account corresponded to a real and identifiable person; but this information was not revealed to other users. From the point of view of users there would have been little apparent change. The users could not see other users’ real names, so this is not a real name environment. We describe it as an environment of stable pseudonyms.
Finally, in June 2014 HuffPo outsourced commenting to Facebook altogether, resulting in a ‘real name’ environment. In this current phase, comments appear below the line of the news article under the user’s real name, as well as – depending on a user’s privacy settings – appearing simultaneously on their Facebook page. We call these phases ‘anonymous’, ‘pseudonymous’, and ‘real name’. There is, of course, more going on than simply the identifiability of users, but we think the terms capture the main markers of change.
What did we find?
You might expect that participation would drop off through each of the changes. And it did.
Patterns of interaction also changed in an interesting way. For the anonymous and pseudonymous phases the majority of commenting had taken place between users, often in the form of spidering conversations, that is, comments on comments and multiple sub-threads. Real name commenting through Facebook exhibits a different pattern: now a much larger proportion of posts were made in direct response to the article with less engagement between users. For the anonymous and pseudonymous commenting regimes, 70% and 68% of comments respectively were replies. For Facebook’s real name environment the percentage is only 54.
In the Facebook phase, commenters were also less likely to engage repeatedly with the platform: in the anonymous phase, the median comment number per user was 4; the figure fell to 3 in the pseudonymous phase, before plummeting to 1 under Facebook commenting. Less than 1% of commenters clocked up 100 comments, compared to 6% in the pseudonymous phase.
Only ⅓ of Real Name commenters commented on more than one politics article in the data set, compared to 60% for both the other periods. In the Facebook phase more people comment only infrequently, and there is a much smaller core of hyperactive commenters. And while this core is too small to allow generalisations beyond the observed data, members of the third phase do not live up to the high standards of users active under the first two phases.
We read this as a shift to a more declarative style of communication, an environment where there is less substantive engagement between individuals. Users are less likely to engage with others, and less likely to engage in long, repeated exchanges of views.
You might also expect users to become more civil. And they did. But then they didn’t. Let us explain.
The move from anonymity to pseudonymity saw a large (40%) reduction in offensive language. This trajectory continued after the second change as well, with a further 10% reduction in the incidence of offensive language. Despite the drop in potentially offensive words, however, insults directed at individuals are on the rise: among replies that use the pronoun ‘you’ at least once, in 1.2% of real name cases ‘you’ is followed by an insult, a 40% rise on pseudonymous comments.
For all three periods we observe a learning effect: if we split each user’s commenting history into an early and a late part, on average the late part contains fewer offensive words than the early one. The effect is not large: in each of the three periods it holds for about 53% of commenters.
In addition to the learning effect, there is a cleansing effect: some commenters choose to leave or are banned for poor behaviour. As a result, commenters with very high comment counts will tend also to be among the best behaved. This effect, whereby the most active users also produce the highest quality contributions, is marked for both the anonymous and pseudonymous phase (we discuss this result in more detail here). The effect for the Facebook environment, though, appears much less pronounced.
Both in terms of general negativity and outright insulting language directed at women, we find, in a small sample of qualitatively graded material, that the pseudonymous phase has roughly half the level of abusive behaviour.
Within a sample of manually coded comments, we found that of root comments including the word ‘she’ 65% were negative under FB commenting, compared to 37% under the pseudonymous phase. 25% were outright insulting, compared to 11% under the pseudonymous phase.
We tested this quantitatively as well, by counting the number of comments that took the format [gendered pronoun] ‘is a’ followed by any word from a list of insults. Though the proportions captured by this test are much smaller (see Figure 2), they point in the same direction: people were the least likely to be rude about women in the pseudonymous environment.
It is perhaps worth noting that this is not simply a ‘Hillary’ effect. We find strong increases in abuse directed against all major female politicians, including Sarah Palin, Michele Bachmann, Elizabeth Warren, and Nancy Pelosi. The proportion of insults within comments on articles about female politicians also doubles between the pseudonymous and real-name phases.
Why everything should not be connected
The question of how people talk about politics online is important. Many people in America today read and discuss the news online, and they increasingly do so through social media platforms. A lot of attention has been paid to fears that this will create ‘filter bubbles’.
But beyond just exposure to information, we also need to consider how, and with whom, people discuss it. And perhaps just as importantly, we have to attend to the different effects of online institutional designs within which this communication takes place.
We don’t have complete answers to the question of deliberative quality under the different commenting architectures we have studied. Deliberative quality is hard to measure. But respectful engagement is one part of it, and looking at the degree and nature of insulting and abusive language, our results suggest that the best commenting behaviour was found in the pseudonymous phase.
Furthermore, while in the anonymous and pseudonymous phase we saw something of a learning process, in which bad behaviour was filtered out as users engaged more in the forum, this relationship breaks down in the Facebook phase. Even though comments still appear below the line on the Huffington Post, it seems that the center of gravity has shifted from single forum of the HuffPo site to the multiple commenting communities of Facebook.
One possible explanation for what is going on is that attempts to enforce community norms and standards are either fewer, less successful, or both. For whatever reason, the new system has not formed a highly committed community.
On a more speculative note, we might say that giving people the freedom to maintain stable pseudonyms without being connected or integrated to wider social networks may be of great value for the willingness to engage in the risky and uncomfortable business of arguing about matters of common concern with people who have very different opinions, perspectives and experiences. Keeping a public and pseudonymous platform but demanding that users authenticate themselves, may create better conditions for discussion than the real name environment of social media.
Anonymous commenting poses real problems, and moderation is crucial but costly. But there are many ways of designing spaces for online discussion, and more attention should be paid to the distinctive potentials of pseudonymity. News media should be wary of ditching their in-house commenting systems in favour of partnerships with entities such as Facebook promising real-name architectures. Perhaps not everything should be connected.
Alfred Moore is a research fellow at Cambridge University, at the Centre for Research in Arts, Humanities and Social Sciences, working on political theory, deliberative democracy, and the politics of expertise. He has taught philosophy at University College Cork, and was a Marie Curie Research Fellow at the University of British Columbia, and a Democracy Fellow at the Ash Center in the Fall of 2012. His current work is funded by the Leverhulme Trust under the project ‘Conspiracy and Democracy: History, Political Theory, Internet.’