Who Uses Bots? A Statistical Analysis of Bot Usage in Moderation Teams - Benjamin Mako Hill
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
Who Uses Bots? A Statistical Analysis
of Bot Usage in Moderation Teams
Charles Kiene Benjamin Mako Hill Abstract
University of Washington University of Washington Adopting new technology is challenging for volunteer mod-
Seattle, WA, USA Seattle, WA, USA eration teams of online communities. Challenges are aggra-
ckiene@uw.edu.edu makohill@uw.edu vated when communities increase in size. In a prior qual-
itative study, Kiene et al. found evidence that moderator
teams adapted to challenges by relying on their experience
in other technological platforms to guide the creation and
adoption of innovative custom moderation “bots.” In this
study, we test three hypotheses on the social correlates of
user innovated bot usage drawn from a previous qualitative
study. We find strong evidence of the proposed relation-
ship between community size and the use of user innovated
bots. Although previous work suggests that smaller teams
of moderators will be more likely to use these bots and that
users with experience moderating in the previous platform
will be more likely to do so, we find little evidence in support
of either proposition.
Author Keywords
moderation; bots; online communities; population size; so-
cial computing; chat; computer-mediated communication;
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed Reddit; Discord.
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored. CCS Concepts
For all other uses, contact the owner/author(s).
CHI ’20 Extended Abstracts, April 25–30, 2020, Honolulu, HI, USA.
•Human-centered computing ! Empirical studies in
© 2020 Copyright is held by the author/owner(s). collaborative and social computing; •Information sys-
ACM ISBN 978-1-4503-6819-3/20/04.
DOI: https://doi.org/10.1145/3334480.3382960
tems ! Chat; Web interfaces;
LBW123, Page 1
CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
Introduction Background
Groups that adopt new organizational technology face Increasing population size presents unique challenges for
unanticipated problems resulting from a disconnect be- governing online communities [1, 8, 13]. For online com-
tween technology designers’ intentions and the group’s munities managed by volunteer moderators, this challenge
technological frames [9, 10]. Other work has found that end becomes even more problematic as volunteers often lack
user programming toolkits not only allow groups to develop the resources to scale their work in response to increasing
innovative custom solutions to problems [14, 15], but that membership and activity [5, 7]. Volunteer moderators are
these solutions are often modeled after tools that groups tasked with shaping and maintaining norms in their commu-
used in prior settings. A 2019 study by Kiene et al. found nities [3, 12]. In many cases, moderation teams adopt user
that moderation teams for online communities that had innovated tools in the form of “bots” to better scale their
adopted a new technological platform felt challenged as a work in response to community growth [2, 4, 6, 7, 11].
result of the new platform’s different affordances, limited
In Kiene et al. ’s 2019 study, the authors interviewed mem-
built-in moderation tools, and problems from increasing
bers of volunteer moderation teams from the social media
community size [7]. These moderation teams adapted to
platform Reddit, a social platform that hosts millions of on-
technological change and problems caused by increasing
line communities called “subreddits” (Figure 1)[7]. Each
community size by utilizing the new platform’s public ap-
team had expanded into Discord, a synchronous chat appli-
plication programming interface (API) to develop, use, and
cation that also hosts millions of online communities called
share custom moderation “bots” that resembled moderation
“servers” (Figure 2) as a second site for community inter-
tools they used on the previous platform.
action. Moderators expressed difficulties in managing their
In this paper, we provide a quantitative test of three hy- communities as a result of inadequate moderation tools
potheses drawn from Kiene et al. [7]. First, we test the on Discord as well as difficulties in scaling their work in re-
Figure 1: A screenshot of the user
proposition that membership size is positively associated sponse to increasing community growth. Moderation teams
interface for a subreddit
community. with the deployment of user created moderation tools. Sec- adapted to these challenges through end user programming
ond, we test the proposition that smaller teams with fewer innovations in the form of bots that facilitated moderation
resources to devote to moderation will be more likely to work. They also found that the features of the bots most fre-
adopt user innovated moderation bots. Finally, we examine quently mentioned in interviews resembled features of mod-
whether teams with experience moderating in Reddit will be eration tools built into Reddit: AutoModerator, Mod Logs,
more likely to adopt user innovated moderation bots. To test and Mod Mail [7].
these theories, we randomly sample 300 online communi-
A bot with an “AutoModerator” (or “automod”) like feature is
ties on the chat platform Discord associated with communi-
depicted in Figure 3. This tool allows teams to scale content
ties on the social media platform Reddit. We found support
moderation by automation. Moderators use the tool to se-
for Kiene et al.’s claim that larger Discord communities will
lect which words, phrases, or even URLs should be filtered
be more likely to adopt user created moderation tools and
from their community; the bot will constantly scan the Dis-
Figure 2: A screenshot of the user no evidence in support of the other propositions.
interface for the Discord application
cord community’s chat channels and automatically delete
and a Discord community. messages according to parameters set by the moderation
LBW123, Page 2
CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
team. Figure 4 depicts a user-created bot that performs a Methods
“Mod Logs” like function. The “mod logs” functionality au- Data
tomatically records and stores information about commu- Data for this study were collected over a three stage pro-
nity members’ interactions with the moderation team, such cess from October to December of 2019. Following Kiene
as when they’re warned for breaking a rule. Finally, some et al. we were interested in communities on Discord that
bots have a “Mod Mail” like feature, depicted in Figure 5, were connected to communities on Reddit. To build a sam-
which aggregates messages from community members ple of these communities, we began with a complete list of
to the moderation team in a centralized location, usually 1,082,444 subreddit communities on Reddit published in
a private chat channel that only the moderators can see. April 20, 2018.1 Next, we gathered data from each of these
Figure 3: A screenshot of a Kiene et al. found that all three of these tools were explicitly subreddits using the Reddit API. This data included each
Discord bot’s automod tools. and closely modeled after Reddit’s moderation tools and all subreddit community’s total subscriber count and text in
three use names that are similar to systems in Reddit. the “side bar”—a column of text on each subreddit’s page.
Representatives of moderation teams from smaller Dis- These “side bars” often contain invite links to any Discord
cord communities interviewed by Kiene et al. reported feel- community associated with the subreddit. Discord com-
ing less of a need to adopt user innovated tools to manage munities can only be accessed through these invite links.
their communities[7]. Therefore, we expect that moderation Using regular expression searches as a method of captur-
teams of Discord-Reddit communities with more members ing invite links to Discord servers, we filtered our initial data
will be more likely to adopt bots that facilitate automod, mod set for subreddit side bars that contained any URL contain-
logs, and mod mail. This leads us to our first hypothesis: ing discord.gg/ or discordapp.com/invite/. This new
(H1) Moderation teams of communities with larger mem- data set contained 8,296 observations.
Figure 4: A screenshot of the
Discord bot “Dyno” and its mod berships will be more likely to adopt bots with moderation We observed that the distribution of total subscribers across
logs feature. features. subreddits in this data set was extremely right-skewed—i.e.,
Kiene et al. also found that moderation teams reported there were many communities with very few subscribers.
turning to bots in order to deal with the problem of limited Because testing H1 requires variation in membership size,
volunteers[7]. In this sense, it stands to reason that larger we stratified this data set into 10 bins based on the num-
teams will be less likely to adopt user innovated bots. This ber of subscribers on an exponential scale. From each bin,
leads us to our second hypothesis: (H2) Smaller modera- we randomly sampled 30 subreddits with working Discord
tion teams will be more likely to adopt bots with moderation invite links. This resulted in a final data set of 300 Discord
features. communities associated with communities on Reddit with a
range of membership sizes. We followed invite links to visit
Finally, Kiene et al. found that moderation teams relied each of these Discord servers and collected data on total
heavily on their technological frames drawn from their ex- community members, members online, number of moder-
Figure 5: A screenshot of a perience in Reddit solving problems with Discord’s API[7].
Discord bot’s mod mail utility. This supports our final hypothesis: (H3) Moderation teams 1
https://www.reddit.com/r/ListOfSubreddits/comments/8drbn3/
with subreddit moderators will be more likely to adopt bots i_created_a_txt_list_of_all_subreddits_1082444_of/ (Archived at
with moderation features. https://perma.cc/FDN2-TDMT)
LBW123, Page 3
CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
Var Min Max ators online, number of Reddit moderators, and the names bot with mod logs feature. Finally, mod mail was recorded
Members 2 180,944 and number of bots in each Discord community manually in as 1 for communities with any bot with a mod mail feature
Mods 0 20 a spreadsheet. or which was explicitly labeled with some variation of “mod
Bots 0 20 mail bot.” We also constructed a dummy variable Any that
Kiene et al. 2019 found that many of the bots used by Dis-
was set to 1 if a community had any of the three bot vari-
cord moderators are published openly and reused across a
Var µs ˙s ables.
range of communities[7]. After recording the names of each
Members 6,731 20,194 bot from each Discord server in our sample, we constructed Independent variables
Mods 3.31 3.26 a data set with details of each bot observed in our second To test H1 about scale, our key question predictor is mem-
Bots 3.50 3.52 data set by conducting Internet searches of the bot’s name bers, which we measure as each Discord community’s total
for its documentation and recording whether the bot has members as published on the Discord server. We note that
Var µw ˙w any of the three moderation features referenced in the prior this captures the total offline and online community mem-
study: automod, mod logs, or mod mail. We then merged bers. Although we also measured the number of people
Members 5,716 19,996
this bot-level data set with the community-level data set to online, preliminary analyses revealed a strong, positive cor-
Mods 2.75 2.97
identify which of the Discord communities had a bot that relation between total community members and members
Bots 3.08 3.45
could perform automod, mod logs, or mod mail functions. online (Spearman’s ˆ = 0.96). To avoid colinearity in our
Additionally, we manually checked each Discord’s commu- models, we use only total community members because
Table 1: Summary statistics for the
nity’s posted “rules” and information channels to determine we believe it’s a more accurate measure of community size
Discord communities in our
if any bots we couldn’t find with Internet searchers were than the count of online members at the time of observation
analysis. s subscripts indicate
being explicitly referenced as moderation tools. In some which could vary by time of day or day of the week.
sample statistics and w subscripts
indicate statistics re-weighted to cases, moderators would tell community members to mes-
Community members on Discord can be assigned cus-
compensate for over-sampling. sage one of their bots to get in touch with the moderation
tom “roles” that signify a members’ role in the community.
(n = 300) team—information that reveal the presence of a mod mail
In most cases, moderators are given some variation of a
bot. In total, we recorded 366 unique bot names, many of
“Moderator” role by the community’s admins. To test H2
which were shared across multiple Discord communities.
about moderation team size, we construct a variable mods
Of these, 32 bots have an automod feature, 45 bots have
which we measure as the number of Discord members with
a mod logs feature, and 11 have a mod mail feature. Sum-
any variation of a “moderator” or “admin” role.
mary statistics for the final sample of projects used in our
analysis can be found in Table 1. In Kiene et al. 2019, the authors reported observing a num-
ber of Discord communities that had moderators who were
Dependent variables
explicitly given tags for the roles of “subreddit modera-
We constructed four dummy variables to capture variation
tors.” To test H3 about moderator experience in Reddit,
in the presence or absence of user-innovated bots drawing
we include the variable redditmods which we measure as
on technological frames from Reddit. First, automod was
a dummy variable set to 1 if a Discord community had mod-
marked as 1 for a community with any bot with an AutoMod-
erators who were explicitly given the role “subreddit moder-
eration feature such as a customizable word or URL filter.
ators” and 0 otherwise. In our data set, we observed a total
Mod logs was recorded as 1 for any observation that had a
LBW123, Page 4CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
of 60 Discord communities that had at least one moderator with 2 total members to have a 20% probability of having
M1 M2
wit a “subreddit moderator“ tag. any user innovated moderation bot while a community with
(Intercept) -1.80 -3.09
939 members (the sample median) to have a 47% probabil-
(0.35) (0.57) Analytic strategy
ity of having any user innovated moderation bot. A commu-
ln(members) 0.20 -0.01 Our analytic strategy employs a logistic regression on the
nity with 13,568 members (90th percentile of our sample)
(0.06) (0.10) probability that any of three moderation tools would appear
would have a 60% probability of having any user innovated
ln(mods) 0.29 0.75 on Discord communities in relation to community size (M1)
moderation bot.
(0.22) (0.36) or that each of the tools would appear individually (M2, M3,
redditmods -0.22 0.30 and M4). In preliminary explorations of our data, we no- In terms of H2, our models provided little in the way of com-
(0.38) (0.52) ticed extreme right skew for the distributions of members pelling evidence in favor of the hypothesis. Although we
Deviance 384.79 190.53 and mods. As a result, we use the natural log of each co- consistently estimate a positive relationship between the
M3 M4 variate in our models. Finally, we apply weights in each of number of moderators and the probability of adopting a
(Intercept) -1.66 -10.92 our analyses to compensate for the oversampling on large user innovated moderation tool, we only estimate a sta-
(0.35) (2.16) communities so that our results reflect the average effects tistically significant relationship between moderation team
ln(members) 0.18 0.66 among any Discord server in the population from which we size and the probability of adopting an automod bot (M2).
(0.06) (0.25) sampled. Our first model (M1) estimates the probability that The relationship is opposite in sign to our prediction.
ln(mods) 0.23 1.33 any of the three moderation tools are adopted through bots,
Our results for H3 are a consistent null finding. We found no
(0.22) (0.73) and the other 3 models test the probability of encountering
statistically significant relationships between our measure of
redditmods -0.35 0.61 each moderation tool in turn. Each model took the same
redditmods and the likelihood of adopting tools across any
(0.38) (0.77) form with the response variable Y changing depending on
of our models.
Deviance 388.66 58.50 which moderation tool we were testing for:
Num. obs. 300 300 Discussion
logit(Y ) = 0 + 1 log(members) + 2 log(mods + 1) +
p < 0.001,
p < 0.01, p < 0.05
Our statistical analyses reported in Table 2 and illustrated
3 redditmods in Figure 6 provide evidence in support of H1. On average,
Table 2: Statistical models and moderation teams are more likely to adopt user innovated
estimates for our four models. M1 Results mod logs or mod mail tools in larger communities. Surpris-
is a logistic regression model of the Results of each statistical test of our models are illustrated ingly, there was virtually no relationship between commu-
probability that a Discord in Table 2. Three of our four fitted models show support for nity size and the adoption of a user created automodera-
community will adopt any user H1. We found a statistically significant relationship between tion tool. This finding was surprising in light of Seering et
innovated moderation bot. M2 - M4 size and the adoption of any moderation bot (M1) as well as al.’s [11] finding where the authors found that larger Twitch
are logistic regression models of for a mod logs bot (M3) and a mod mail bot (M4). We did streams resulted in more automated moderation activity
the probability that a Discord not find a statistically significant relationship between size from bots. This could be explained by the fact that while
community will adopt the three
and the likelihood of using a bot with the automod feature. Twitch streams only allow one chat channel for commu-
moderation tools: automod (M2),
Figure 6 depicts predicted values from all four models for nity interactions, Discord affords the creation of up to 500
mod logs (M3), and mod mail (M4).
prototypical Discord communities with varying community different chat channels in a single Discord community. Fur-
sizes holding the other variables at their sample median val- thermore, these findings might be interpreted as suggesting
ues. For example, We would expect a Discord community
LBW123, Page 5CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
0.8 that increasing community size leads to greater challenges ally, the measure for the number of moderators available
0.6 related to logging member information and less problems to us was the number of members with any variation of a
M1: Any
0.4
for scaling content moderation. “Moderator” role online at the time that we visited each Dis-
cord server. We might expect to find different moderators
0.2 Although the effect of community size on the probability
online at other times. Unfortunately, Discord, unlike Reddit,
0.0 of adopting a mod mail bot was small, this might be ex-
does not have an API that lets users index Discord com-
0.8
plained by the fact that most mod mail bots we observed
munities for the number of moderators. Our measure for
M2: Automod
in our sample were custom-built and self-hosted on private
0.6 whether subreddit moderators are present in the Discord
servers. On the other hand, almost all bots with mod log
0.4 community is also limited in that moderation teams for Dis-
features in our sample were available as public services
0.2 cord communities are not required to label or make “sub-
that did not require community moderators to self-host the
reddit mod” roles. It is possible that subreddit moderators
0.0 bot on their own servers. This may explain the relatively
could have been present as Discord moderators despite not
0.8 widespread use of these bots Discord communities in our
also being given a “subreddit mod” role. Noisiness in this
M3: Modlog
0.6 sample. The positive association of the parameter esti-
measure may contribute to our null results for H3.
0.4 mates in M2 is opposite in sign for the effect we expected
in H2. After controlling for the log-linear effect of commu- Conclusion
0.2
nity size, larger moderation teams are actually associated Despite its limitations, our study offers several contributions
0.0
with a higher likelihood of adopting user innovated auto- to existing research on user innovation, community mod-
0.8 mod tools. In that ln(members) and ln(mods) are positively eration, and the effects of community size. As precited by
M4: ModMail
0.6 correlated (Pearson’s ˆ = 0.62), we find that larger commu- Kiene et al., we found evidence that online communities are
0.4 nities tend to have more mods. It is possible that our control more likely to adopt user innovated moderation tools when
0.2 for the log-linear effect of membership may be incomplete they operate at scale. Extending previous work, we found
0.0 in ways that are reflected in our estimate of the marginal ef- that the effect of size was strongest for moderation tools
10 1,000 100,000 fect for ln(mods). It is also possible that larger teams have that allowed the systematic logging and indexing of moder-
Discord members more time, technical skills, or organizational slack to devote ation actions like warnings and bans. Our results suggest
to setting up bots. that certain types of work for volunteer moderators, like or-
Figure 6: Predicted probabilities ganizing and managing online communities, may benefit
Limitations
for each model based on a logged from automation at scale more than others.
Our study has a number of important limitations. The sam-
increase in community size. Tick
marks on the x-axis indicate
ple size of 300 is relatively small. This may explain our Acknowledgements
sample observations for community relatively large standard errors and some of our null find- This work was supported by the National Science Founda-
size. Shaded areas represent the ings. Additionally, our dependent variables represent only tion (awards IIS-1617129 and IIS-1617468). Feedback and
95% prediction interval. the presence of user innovated moderation bots in Discord support for this work came from Kirsten Foot, members of
communities but does not guarantee that the features of the the Community Data Science Collective, and from the Uni-
bots are being routinely used. For example, many commu- versity of Washington Department of Communication. The
nities in our sample had multiple bots that performed similar manuscript benefited from excellent feedback from several
functions which might imply that only one is used. Addition- anonymous reviewers at CHI-LBW.
LBW123, Page 6CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
REFERENCES [6] Shagun Jhaver, Iris Birman, Eric Gilbert, and Amy
[1] Brian S. Butler. 2001. Membership size, Bruckman. 2019. Human-machine collaboration for
communication activity, and sustainability: A content regulation: the case of reddit automoderator.
resource-based model of online social structures. ACM Trans. Comput.-Hum. Interact. 26, 5 (July 2019),
Information Systems Research 12, 4 (2001), 346–362. 31:1–31:35. DOI:
DOI: http://dx.doi.org/10.1145/3338243
http://dx.doi.org/10.1287/isre.12.4.346.9703 [7] Charles Kiene, Jialun "Aaron" Jiang, and
[2] Eshwar Chandrasekharan, Chaitrali Gandhi, Benjamin Mako Hill. 2019. Technological frames and
Matthew Wortley Mustelier, and Eric Gilbert. 2019. user innovation: exploring technological change in
Crossmod: A Cross-Community Learning-based community moderation teams. Proceedings of the
System to Assist Reddit Moderators. Proceedings of ACM: Human-Computer Interaction 3, CSCW (Nov.
the ACM on Human-Computer Interaction 3, CSCW 2019), 44:1–44:23. DOI:
(Nov. 2019), 174:1–174:30. DOI: http://dx.doi.org/10.1145/3359203
http://dx.doi.org/10.1145/3359276 [8] Charles Kiene, Andrés Monroy-Hernández, and
[3] Eshwar Chandrasekharan, Mattia Samory, Shagun Benjamin Mako Hill. 2016. Surviving an “Eternal
Jhaver, Hunter Charvat, Amy Bruckman, Cliff Lampe, September”: How an online community managed a
Jacob Eisenstein, and Eric Gilbert. 2018. The surge of newcomers. In Proceedings of the 2016 ACM
internet’s hidden rules: An empirical study of reddit Conference on Human Factors in Computing Systems
norm violations at micro, meso, and macro scales. (CHI ’16). ACM, New York, NY, 1152–1156. DOI:
Proc. ACM Hum.-Comput. Interact. 2, CSCW (Nov. http://dx.doi.org/10.1145/2858036.2858356
2018), 32:1–32:25. DOI: [9] Wanda J. Orlikowski. 1992. Learning from notes:
http://dx.doi.org/10.1145/3274301 Organizational issues in groupware implementation. In
[4] R. Stuart Geiger. 2014. Bots, bespoke, code and the Proceedings of the 1992 ACM Conference on
materiality of software platforms. Information, Computer-supported Cooperative Work (CSCW ’92).
Communication & Society 17, 3 (March 2014), ACM, New York, NY, USA, 362–369. DOI:
342–356. DOI: http://dx.doi.org/10.1145/143457.143549
http://dx.doi.org/10.1080/1369118X.2013.873069 event-place: Toronto, Ontario, Canada.
[5] Tarleton Gillespie. 2018. Custodians of the Internet: [10] Wanda J. Orlikowski and Debra C. Gash. 1994.
platforms, content moderation, and the hidden Technological frames: Making sense of information
decisions that shape social media. Yale University technology in organizations. ACM Trans. Inf. Syst. 12,
Press, New Haven. OCLC: on1005113962. 2 (April 1994), 174–207. DOI:
http://dx.doi.org/10.1145/196734.196745
LBW123, Page 7CHI 2020 Late-Breaking Work CHI 2020, April 25–30, 2020, Honolulu, HI, USA
[11] Joseph Seering, Juan Pablo Flores, Saiph Savage, [13] Aaron Shaw and Benjamin Mako Hill. 2014.
and Jessica Hammer. 2018. The social roles of bots: Laboratories of oligarchy? How the iron law extends to
Evaluating impact of bots on discussions in online peer production. Journal of Communication 64, 2
communities. Proc. ACM Hum.-Comput. Interact. 2, (2014), 215–238. DOI:
CSCW (Nov. 2018), 157:1–157:29. DOI: http://dx.doi.org/10.1111/jcom.12082
http://dx.doi.org/10.1145/3274426
[14] Eric von Hippel. 2016. Free innovation (1 edition ed.).
[12] Joseph Seering, Robert Kraut, and Laura Dabbish. The MIT Press, Cambridge, MA.
2017. Shaping Pro and Anti-Social Behavior on Twitch
Through Moderation and Example-Setting. In [15] Eric von Hippel and Ralph Katz. 2002. Shifting
Proceedings of the 2017 ACM Conference on innovation to users via toolkits. Management Science
Computer Supported Cooperative Work and Social 48, 7 (2002), 821–833.
Computing (CSCW ’17). ACM, New York, NY, USA, http://www.jstor.org/stable/822693
111–125. DOI:
http://dx.doi.org/10.1145/2998181.2998277
LBW123, Page 8You can also read
