Anyone Can Become a Troll: Causes of Trolling Behavior in Online Discussions

INTRODUCTION

• Online antisocial behavior causes emotional distress and can lead to offline harassment and violence; it is prevalent across web sites, with over one in five CNN.com comments removed.The introduction also reports that 40% of internet users had experienced online harassment.Source: INTRODUCTION · p. 1
• Over 16 million CNN.com posts showed that one in four abuse-flagged posts came from users with no prior record, alongside mood and exposure effects.Trolling propensity tracked population-level mood shifts; exposure effects varied with the volume and ordering of prior troll posts, and mood showed persistence and temporal decay.Source: INTRODUCTION · p. 1
• A logistic regression model achieved AUC=0.78, with discussion context and recent posting-history mood explaining trolling better than an innate-trait account.Context was whether prior discussion posts were flagged; mood was reflected in whether the user’s recent posts elsewhere were flagged.Source: INTRODUCTION · p. 2

Antisocial behavior in online discussions

• The paper defines trolling to include flaming, griefing, swearing, and personal attacks outside community guidelines, using manual coding experimentally and abuse flags in longitudinal analysis.Prior literature offers several definitions, including disruption, pleasure in upsetting others, and broadly negatively marked behavior.Source: Antisocial behavior in online… · p. 2

BACKGROUND

• The background review frames whether situational factors can trigger trolling and how it spreads through a community as open questions.The paper combines a field experiment to isolate causal mechanisms with observational analysis to examine them in the wild.Source: INTRODUCTION · p. 1, Antisocial behavior in online… · p. 2, BACKGROUND · p. 2

Antisocial behavior in online discussions

• Although prior accounts portray trolls as a small, predisposed group, the authors ask whether ordinary users can be situationally influenced to troll.Online disinhibition and anonymity can reduce accountability and comment quality.Source: Antisocial behavior in online… · p. 2

Causes of antisocial behavior

• Earlier accounts proposed boredom, fun, or venting as motivations for antisocial behavior, but were largely qualitative and non-causal.Whether these motivations apply to the general population remained largely unknown, motivating the focus on mood and discussion context.Source: Causes of antisocial behavior · p. 2

Mood

• Mood-linked circumstances may shape trolling: the authors expect more trolling when mood worsens and less susceptibility after time to cool down.They point to mood changes by time of day or day of week and to evidence that “timeouts” can reduce aggression.Source: Causes of antisocial behavior · p. 3
• Negative mood is hypothesized to increase trolling, though its effects on commentary may be mixed.Negative mood may impair self-regulation and worsen impressions of others, while positive and negative affect have both been linked to stereotyping, which may trigger trolling. Negative moods may also promote more thoughtful commentary, but the authors expect negative effects in social contexts to outweigh such factors.Source: Causes of antisocial behavior · pp. 2–3

Discussion context

• Discussion norms and earlier contributions may shape later comments, but users can also respond to trolling with sympathy, apologies, or jokes.The authors note that such responses are rarer and extend prior work on comment cascades to trolling in experimental and observational settings.Source: Causes of antisocial behavior · p. 3
• The authors hypothesize that discussion context, especially prior troll posts by other users, affects a user’s likelihood of trolling.They also expect the topic of a news discussion to play a role in subsequent trolling.Source: Causes of antisocial behavior · p. 3

Influence and antisocial behavior

• Prior work links social influence, emotional and behavioral transfer, and responses to downvotes to people repeating or spreading others’ behavior [17].Examples include herding toward others’ actions and downvoted users later downvoting others or posting content that is further downvoted.Source: Influence and antisocial behavior · p. 3
• The Broken Windows hypothesis suggests that comments in poor taste can invite worse comments and that normalized antisocial behavior can perpetuate itself.The analogy is that an unfixed broken window can create a perception of unruliness.Source: Influence and antisocial behavior · p. 3
• Negative entities are described as more contagious than positive ones, while bad impressions form faster and resist disconfirmation, leading the authors to expect antisocial behavior to be influential and persistent.This reasoning motivates the expectation that trolling can spread from user to user (H3).Source: Influence and antisocial behavior · p. 3
• The authors test H3 by tracing CNN.com discussion evolution and developing an overall model of how trolling might spread between users.They first test H1 and H2 in a controlled experiment, then verify and extend those results using CNN.com discussions.Source: Influence and antisocial behavior · p. 3

EXPERIMENT: MOOD AND DISCUSSION CONTEXT

• The two-by-two between-subjects design varied positive versus negative mood and benign versus troll-like initial discussion posts.Participants were assigned round robin to the four conditions.Source: EXPERIMENT: MOOD AND DISCUSSION CONTEXT · p. 3, Figure 1 · p. 4
• Discussion quality was assessed through expert labels of trolling and LIWC negative-affect scores; the experts were blind to condition and resolved disagreements by discussion.Trolling labels covered posts that were offensive, irrelevant, or designed to elicit anger, whether intentional or not; Vader gave similar affect results.Source: EXPERIMENT: MOOD AND DISCUSSION CONTEXT · p. 3, Experimental Setup · p. 5

Experimental Setup

• The AMT experiment had a quiz followed by a discussion, restricted participation to US residents, and compensated participants $2.00.Participants were not told the study’s purpose until debriefing, and the protocol was reviewed under IRB Protocol #32738.Source: Experimental Setup · p. 3
• The mood manipulation used a timed 15-question quiz, with harder questions and expected negative feedback in NEGMOOD versus easier questions and expected positive feedback in POSMOOD.Both quizzes had the same question-type composition and order, a five-minute limit, and feedback compared with a fixed average of eight correct answers.Source: Experimental Setup · p. 4
• After the quiz, participants completed 65 Likert-scale mood questions based on the Profile of Mood States questionnaire as a manipulation check.POMS measures mood on six axes, including anger and fatigue.Source: Experimental Setup · p. 4
• In the discussion, participants saw either three initial troll comments in NEGCONTEXT or innocuous comments in POSCONTEXT, then could comment, reply, and vote.The simulated news-site interface used an abridged article about the 2016 Democratic primaries.Source: Experimental Setup · p. 4
• Both NEGMOOD and NEGCONTEXT significantly increased trolling probability; trolling and negative-affect proportions were lowest in POSMOOD/POSCONTEXT and highest—and almost double—in NEGMOOD/NEGCONTEXT.The table caption does not provide the underlying proportions or regression estimates.Source: Table 1 · p. 4, Table 2 · p. 5
• Eight independent universes per condition—32 total—limited dependence on chance breakdowns, while allowing participants within each universe to interact.Participants were randomized among universes within their condition; universes shared seed comments but not participants or interactions.Source: Experimental Setup · p. 5

Limitations

• The experiment required 667 participants to post at least one comment, and payment and pseudonymous usernames may have affected participation incentives or comment quality.Different initial posts and article topics may also elicit different trolling; demographic effects, trolling types, and user motivations remain for future study.Source: Results · pp. 5–6

Manipulation checks

• Mood disturbance was 40.8 in NEGMOOD versus 12.2 in POSMOOD, while initial posts were upvoted 36% in NEGCONTEXT versus 90% in POSCONTEXT.These checks supported that the quiz shifted mood and that the negative-context seed posts were perceived as more troll-like.Source: Results · p. 5

Trolling behavior

• Troll posts peaked at 68% under NEGMOOD and NEGCONTEXT, versus 47%, 49%, and 35% in the other conditions; negative mood raised trolling odds 89% and prior troll posts raised them 68%.Both NEGMOOD and NEGCONTEXT had significant effects (p<0.05); the results remained similar in MCMC and demographic-controlled models.Source: Results · p. 5

Affect measures

• Negative-word affect increased significantly under NEGCONTEXT but not NEGMOOD, and neither condition significantly affected positive affect.The authors note that negative-word measures may miss trolling such as sarcasm or off-topic posting.Source: Results · p. 5

Examples of troll posts

• Troll posts ranged from swearing and personal attacks to veiled insults, sarcasm, and off-topic statements, while non-troll posts were more measured.The contrast held regardless of whether non-troll posts agreed with the article.Source: Results · p. 5

Other results

• Word counts fell from 44 in POSMOOD and POSCONTEXT to 29 in NEGMOOD and NEGCONTEXT, while other-user upvotes fell from 79% to 75%; neither trend was significant.Source: Results · p. 6

Discussion

• NEGCONTEXT may make expressing negative reactions seem more acceptable, while NEGMOOD may heighten perceived negativity and reduce self-inhibition.Source: Results · p. 6

Real-world extension

• The authors report that initial posts significantly affect subsequent trolling in CNN.com discussions and that article topic also mediates trolling behavior.Their subsequent analyses examine daily mood patterns, mood persistence across discussions, and effects of troll-post volume and ordering.Source: Results · p. 6

• 26% of flagged posts came from users with no prior flagging record, while 40% of prolific users with flagged posts had fewer than 3.5% of their posts flagged.These observations suggest ordinary users account for a significant amount of trolling behavior.Source: DATA: INTRODUCTION · p. 6
• The Disqus trace covers December 2012–August 2013: 865,248 users, 16,470 discussions, and 16,500,603 posts; 571,662 posts were flagged and 3,801,774 were deleted.The user count includes 20,197 banned users; 3.5% of posts were flagged and 23% deleted.Source: DATA: INTRODUCTION · p. 6
• The analysis distinguishes discussions following a news article from sub-discussions comprising a top-level post and its replies, since replies entail responding to prior posts.Discussions may span thousands of posts, where users may not read earlier responses.Source: DATA: INTRODUCTION · p. 6
• Expert labels of 500 posts yielded flag precision of 0.66 and recall of 0.94, supporting flags as a primary trolling measure despite some missed troll posts.Moderator deletions, negative affect, and downvotes are less suitable as primary measures for the stated limitations.Source: DATA: INTRODUCTION · p. 6
• Flagged posts had more negative-affect words (3.7% vs. 3.4%, Cohen’s d=0.06, t=40, p<0.001), downvotes (58% vs. 30% of votes, d=0.76, t=531, p<0.001), and moderator deletions (79% vs. 21% of posts, d=1.4, t=1050, p<0.001).These signals are correlated with flags and complement their use as the primary estimate of trolling behavior.Source: DATA: INTRODUCTION · pp. 6–7

DATA: UNDERSTANDING MOOD

• The authors use time of day and day of week as correlates of mood because seasonality influences mood.They examine whether trolling behavior changes across those time periods.Source: DATA: UNDERSTANDING MOOD · p. 7
• Because aggression can linger after an unpleasant event, the authors study whether trolling persists as users participate in multiple discussions.The analysis uses large-scale longitudinal observational data to verify and expand on the earlier experimental finding that bad mood increases trolling.Source: DATA: UNDERSTANDING MOOD · p. 7

Happy in the day, sad at night

• Flagged posts, negative affect, and downvotes were lowest in the morning and highest in the evening, and all three measures peaked on Monday.These aggregate patterns align with times when mood is likely worse.Source: Happy in the day, sad at night · p. 7
• Within the same users, flagged posts rose from 4.1% to 4.3%, negative affect from 3.3% to 3.4%, and downvotes from 20.6% to 21.4% between morning and late night.Posts were also more negative on Monday than Friday (d≥0.02, t>2.5, p<0.05), with similar results when limited to US and World news.Source: Happy in the day, sad at night · p. 7

Anger begets more anger

• 4.6% vs. 2.1%: users with a flagged post in a prior discussion were twice as likely to have their next post in a different discussion flagged.The comparison matched users on prior post counts and excluded users with any prior flagged posts; similar results held with stricter history requirements.Source: Anger begets more anger · p. 7, Figure 2 · p. 8
• The shorter the interval between a user’s flagged post and their next post in an unrelated discussion, the more likely the next post was also flagged.This pattern suggests negative mood may persist for some time.Source: Figure 2 · p. 8
• 2.2% vs. 1.7%: users exposed to a flagged post in a prior discussion were more likely to have a flagged post in a later discussion, though the effect was weaker.Users had no prior flagged posts, and the exposure occurred in a different sub-discussion from the one the user participated in.Source: Anger begets more anger · p. 8
• Both prior trolling and exposure to trolling can predict later trolling among users without prior flagged posts, consistent with negative mood and trolling norms carrying across discussions.The effects are unlikely to arise simply because some users were generally trolls, since the analyzed users had no prior flagged posts.Source: Anger begets more anger · p. 8

Time heals all wounds

• Flagging was most likely when a user posted within five minutes of their own flagged post, then gradually decreased as more time passed, including after ten minutes.The next post was in a new discussion with different other users; elapsed time was examined as a possible proxy for distance from the negative mood associated with the previous flagged post.Source: Time heals all wounds · p. 8
• The authors note that users with better impulse control may wait longer when angry, so the observed time pattern does not isolate the effect of waiting; it lends credence to forum rate-limiting.Source: Time heals all wounds · p. 8

“FirST!!1”

• The initial-post comparison matched discussions on article topic, posting weekday, and popularity, and excluded users most directly connected to the initial post or with prior flagged or deleted posts.Discussions had at least 20 posts; propensity score matching was used to create the pairs.Source: “FirST!!1” · p. 8
• 3.1% vs. 1.7% of other users’ subsequent posts were flagged after an initial flagged post versus an unflagged one, and the difference persisted in the discussion’s second half (2.1% vs. 1.3%).The analysis excluded the initial poster, users who replied to that post, and users with prior flagged or deleted posts in other discussions.Source: “FirST!!1” · p. 8

DATA: UNDERSTANDING DISCUSSION CONTEXT

• 7.1% vs. 1.7% of subsequent posts were flagged when the first three posts were all flagged versus none, while flagging also varied with prior-flag count, flagged-post position, and topic.The gap widened compared with discussions where only the first post was flagged.Source: DATA: UNDERSTANDING DISCUSSION CONTEXT · p. 8, “FirST!!1” · p. 9, Figure 4 · p. 9

“FirST!!1”

• 9.6% vs. 5.9% of later posts were flagged in sub-discussions with a flagged versus unflagged top-level post, even when comparing within the same discussion.This comparison controls for the article being discussed and considers replies within sub-discussions.Source: “FirST!!1” · p. 9

From bad to worse: sequences of trolling

• The analysis examines discussions with at least five posts and distinguishes users new to a sub-discussion from those who have already posted there.This distinction controls for the impact of prior participation in that discussion.Source: From bad to worse: sequences of trolling · p. 9
• For a fifth post by a new sub-discussion user, flagging rose from 2% with no flagged predecessors to 7% with one and 49% with four.The increase was monotonic, and pairwise differences were significant after Holm correction; similar trends held for users who had posted before, who were more likely to be flagged subsequently.Source: From bad to worse: sequences of trolling · p. 9
• When exactly one of the first four posts was flagged, a flagged fourth post was associated with a greater likelihood of a flagged fifth post than a flagged first post.This pattern held for both new and previously participating users (χ2(1)>6.9, p<0.01).Source: From bad to worse: sequences of trolling · p. 9
• The findings support H3: trolling behavior spreads from user to user, and conspicuous troll posts affect whether new discussants troll.The authors also note that participating in a previous discussion containing a flagged post raises the likelihood of future trolling.Source: From bad to worse: sequences of trolling · p. 9

Hot-button issues push users’ buttons?

• Flagging was near 4% in health, justice, showbiz, sport, US, and world sections, versus near 2% in opinion, politics, tech, and travel.The authors suggest controversy may help explain higher rates in health, justice, US, and world, while inter-group conflict may matter in showbiz and sport; politics had a low rate despite its polarizing issues.Source: Hot-button issues push users’ buttons? · p. 9

A MODEL OF HOW TROLLING SPREADS

• Discussion-context features performed best alone (AUC=0.74), while previous posts plus recent history nearly matched all features (0.77 vs. 0.78).User-specific propensity and lifetime trolling history were each moderately predictive (0.66); performance remained 0.75 for users new to a discussion and 0.70 for users without prior reported or deleted posts.Source: Table 3 · p. 10, A MODEL OF HOW TROLLING SPREADS · p. 10

Hot-button issues push users’ buttons?

• The interplay of topic with personal values and group membership in trolling remains future work.Source: Hot-button issues push users’ buttons? · p. 10

Summary

• Bad mood, time of day and day of week, prior troll posts, and more controversial topics all affect trolling, with prior troll posts adding to future likelihood.Mood effects may persist across discussions but diminish with time; trolling also varies by time of day and day of week.Source: Summary · p. 10

A MODEL OF HOW TROLLING SPREADS

• The prediction model represented mood through seasonality and recent posting history, discussion context through prior posts and topic, and innate propensity through user identity and lifetime trolling history.Context features included whether any of the previous five discussion posts were flagged and whether they were written by the same user.Source: A MODEL OF HOW TROLLING SPREADS · p. 10
• The prediction task used 116,026 randomly sampled posts, balanced flagged and unflagged users for 50% random-guess accuracy, and a logistic regression classifier.The analysis included users with prior flagged posts, and a random forest produced empirically similar results.Source: A MODEL OF HOW TROLLING SPREADS · p. 10
• Discussion context predicts later trolling beyond users’ intrinsic propensity, while recent posting history also predicts trolling and may reflect mood carrying over between discussions.The authors interpret past trolling as predictive of future trolling as well.Source: A MODEL OF HOW TROLLING SPREADS · p. 10

DISCUSSION

• Although prior work suggests some users may be innately more likely to troll, the results show ordinary users also troll when mood and discussion context prompt them.The paper frames mood and context as conditions under which trolling is not confined to an antisocial minority.Source: DISCUSSION · p. 10

The spread of negativity

• Flagged-post prevalence rose from 0.03 to 0.04, and the proportion of users with flagged posts rose from 0.09 to 0.12 between dataset halves (p<0.001).Figure 5 shows these upward trends on CNN.com, suggesting trolling may spread and be reinforced.Source: Figure 5 · p. 11, The spread of negativity · p. 11
• On CNN.com, flagged posts and users with flagged posts became more common over time, suggesting trolling behavior may be spreading.The discussion raises the possibility that induced behavior carried over from earlier discussions could cascade.Source: The spread of negativity · pp. 10–11, Figure 5 · p. 11
• One possible explanation is that users who joined later were more susceptible to trolling.The authors note that there may be several explanations for the rising proportions.Source: The spread of negativity · p. 11
• Prior work suggests negative norms can be reinforced and that downvoted users may go on to downvote others [17].These findings support the possibility that negative behavior can persist and permeate a community when unchecked.Source: The spread of negativity · p. 11
• The discussion concludes that negative behavior can persist in and permeate a community when left unchecked.The increasing flagged-post and flagged-user proportions are consistent with this possibility.Source: The spread of negativity · p. 11

Designing better discussion platforms

• Because people may mistake situational trolling for an innate trait, banning users alone may leave ordinary users’ situational trolling unchecked.Some banned users may simply have been having a bad day.Source: Designing better discussion platforms · p. 11
• Inferring mood from recent activity or behavioral traces and applying rate limits, comment retraction, or less frustrating interfaces may curb heated posting.These measures aim to discourage posting in the heat of the moment or reduce frustration that can fuel aggression.Source: Designing better discussion platforms · p. 11
• Hiding troll comments and prioritizing constructive ones may make discussions seem more civil and reduce users’ likelihood of following suit.Source: Designing better discussion platforms · p. 11
• Downvoting can reduce exposure to disliked comments, but may also worsen subsequent comments; selectively showing positive feedback may avoid that feedback loop [17].Source: Designing better discussion platforms · p. 11
• Ethical reminders or pledges can strengthen community norms and increase future moral behavior.Source: Designing better discussion platforms · p. 11

Limitations and future work

• Future work should improve mood signals and model reply structure [3], sentiment changes, and idea flow to clarify mood and context effects.Source: Limitations and future work · p. 11
• Research should distinguish trolling types, anonymity’s effects on misconduct and participation, author intent, and persistent versus situational users [18].Trolling strategies may differ in prevalence and severity; users whose posts were consistently deleted may troll persistently, while those deleted only before bans may troll situationally.Source: Limitations and future work · p. 11

CONCLUSION

• Trolling reflects both innate and situational factors, so addressing conditions that prompt trolling may complement banning users who violate community norms.The paper focuses on situational factors, especially mood and discussion context, and argues that platform design should account for both types of trolling.Source: CONCLUSION · p. 11