💻 computer science

The Power of Social Norms: How Initial Responses to Toxicity Shape Conversations on Twitter

This study analyzes over 187,000 tweets to demonstrate that group dynamics and initial responses to toxicity on Twitter significantly shape subsequent conversation tone, revealing that a larger audience before a toxic post and a toxic first reply both discourage civil responses and foster increasingly abusive interactions.

Original authors: Ana Aleksandric, Mohit Singhal, Anne Groggel, Shirin Nilizadeh

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ana Aleksandric, Mohit Singhal, Anne Groggel, Shirin Nilizadeh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine Twitter (or X) as a massive, bustling town square. Everyone is shouting, sharing stories, and arguing. Usually, it's a lively place, but sometimes, someone starts yelling insults or spreading hate. This is what researchers call "toxicity."

This paper asks a simple but powerful question: When someone starts yelling in the town square, what happens next? Does the crowd jump in to stop it, or do they just watch, or worse, start yelling too?

The researchers looked at nearly 10,000 of these arguments to see how the "crowd dynamics" change the outcome. Here is the breakdown using some everyday analogies:

1. The "Bystander Effect" in the Digital Square

You might know the Bystander Effect from real life: If you see someone drop their groceries in a busy street, you are less likely to help if there are 50 people watching compared to if you are the only one there. Why? Because you think, "Someone else will do it," or "Maybe they don't need help."

The researchers found this happens online, too.

The Analogy: Imagine a toxic tweet is a person tripping and falling in the square.
The Finding: If a huge crowd is already watching the conversation before the fall, fewer people step in to help. The more people watching silently, the less likely anyone is to say, "Hey, that's not okay."
In plain English: The more people are already in the chat before the abuse starts, the fewer people will speak up to defend the victim afterward. The responsibility gets "diluted" among the crowd.

2. The "First Domino" of Tone

The study also looked at what happens immediately after the first insult is thrown. Who replies first?

The Analogy: Think of a conversation like a game of catch. If I throw a ball at you, and you throw it back gently, the game stays friendly. But if I throw a ball at you, and you immediately throw a brick back, the game turns into a fight.
The Finding: The very first reply to a toxic comment sets the "rules" for everyone else.
- Scenario A: If the first person to reply says something kind or calm, the rest of the conversation tends to stay calm. It's like a referee blowing a whistle to stop the fight.
- Scenario B: If the first person to reply throws a brick back (a toxic reply), the whole conversation turns into a brick-throwing match. The "norm" becomes: "It's okay to be mean here."
In plain English: The first response acts as a social cue. If the first response is nice, others follow suit. If the first response is nasty, the whole thread becomes nastier.

3. The "Silent Majority" Problem

The researchers did a deep dive into the actual text of these arguments. They were looking for "Positive Standing Up"—people who politely corrected a misunderstanding, defended the victim, or set boundaries.

The Finding: It's surprisingly rare. In the toxic conversations they studied, only about 20% of the time did someone try to fix the situation or defend the victim.
The Analogy: Imagine a classroom where one kid is bullying another. In 80% of the cases, the other kids either join in, laugh, or stay completely silent. Only 1 out of 5 kids actually raises their hand to say, "Stop that."

4. Why Verified Accounts Get Less Help

Interestingly, they found that if the person being attacked has a "Verified" checkmark (a celebrity or influencer), people are less likely to defend them.

The Analogy: If a regular person gets tripped in the square, people rush to help. If a famous movie star gets tripped, people might think, "They have lawyers," or "They can handle it," or "I don't want to get involved with a celebrity."
The Finding: People are less likely to "stand up" for famous or highly visible accounts, perhaps thinking they don't need the help or that it's too risky to get involved.

The Big Takeaway

The paper concludes that online behavior is contagious.

Toxicity spreads when people see others being mean and no one stopping it.
Kindness spreads when the first person to reply sets a positive tone.

The researchers suggest that to fix this, we need to understand that silence is often interpreted as approval. If we want to stop online bullying, we can't just rely on one person to be a hero; we need to change the "group norm" so that the first person to speak up feels safe and encouraged to set a kind tone, which then encourages everyone else to follow.

In short: The first few seconds of an online argument decide whether the rest of the conversation will be a shouting match or a civil discussion. And if too many people are watching silently, no one feels brave enough to speak up.

1. Problem Statement

Online harassment and toxic language remain pervasive issues on social media platforms like Twitter (X). While previous research has identified factors contributing to toxicity, there is a gap in understanding how group dynamics and social norms influence user behavior in response to toxic content. Specifically, the study investigates:

The Bystander Effect: Does the presence of multiple observers (users participating before a toxic event) diffuse individual responsibility to intervene, leading to fewer non-toxic responses?
Norm Contagion: Does the tone of the first direct reply to a toxic tweet establish a social norm that dictates the trajectory of the entire conversation thread?

The authors aim to determine if passive observation by a group implicitly sanctions toxic behavior and how initial responses can either escalate or de-escalate conflict.

2. Methodology

Data Collection and Preprocessing

Datasets: The study utilizes two random samples of public English tweets:
- Dataset 1: August 14 – September 28, 2021 (5,115 conversations).
- Dataset 2: March 31 – April 23, 2023 (3,992 conversations).
Structure: Conversations are modeled as trees where the "root tweet" initiates the thread.
Filtering Criteria:
- Only conversations containing at least one toxic reply were retained (10,455 initial candidates).
- Conversations where the root author posted the first toxic reply were excluded (to isolate the bystander effect on others).
- Only replies received within 48 hours of the initial tweet were analyzed to ensure temporal consistency.
- Empty tweets (links/images only) and non-English replies (translated to English) were processed.
Final Dataset: 9,107 unique conversations comprising 187,658 tweets from 118,609 unique users.

Toxicity Detection

Tool: Google's Perspective API was used to score toxicity (0–1 scale).
Threshold: A tweet was labeled "toxic" if its score was $\ge$ 0.7.
Preprocessing: Text was cleaned (removed punctuation/URLs, emojis replaced with text) before analysis.

Variables

Independent Variables:
- users_before_toxic: Count of unique users participating before the first toxic comment (proxy for bystander presence).
- reply_toxicity: The toxicity score of the first direct reply to the initial toxic comment.
Dependent Variables:
- unique_users_non_toxic: Count of unique users posting non-toxic replies after the first toxic comment.
- remaining_toxicity: The ratio of toxic replies to total replies in the thread after the first toxic comment (excluding the immediate reply).
Control Variables:
- User Attributes: Follower count, friend count, tweet volume, account age, verification status, profile description length, URL presence, location presence.
- Conversation Structure: Tree width (max tweets at any level) and depth (longest path from root).
- Topics: 19 categories (e.g., politics, sports, entertainment) derived from a RoBERTa-based classifier.

Statistical Analysis

Hypothesis 1 & 2: Tested using Poisson Regression (due to count data distribution) to analyze the relationship between independent variables and the count of non-toxic users.
Hypothesis 3: Tested using Ordinary Least Squares (OLS) Linear Regression to analyze the relationship between initial reply toxicity and the remaining toxicity ratio.
Robustness: Negative Binomial regression was also run to account for over-dispersion, yielding consistent results.

3. Key Contributions and Hypotheses

The study tests three specific hypotheses derived from Social Norm Theory and the Bystander Effect:

H1 (Bystander Effect): The number of users participating before the first toxic reply is negatively associated with the number of users who respond non-toxically.
- Finding: Supported. As the number of pre-toxic participants increases, the likelihood of others intervening with non-toxic comments decreases significantly ( $p < 0.001$ ).
H2 (Norm Setting - Intervention): If the first reply to a toxic comment is non-toxic, more users will post non-toxic replies.
- Finding: Supported. A higher toxicity score in the immediate reply is negatively associated with the number of subsequent non-toxic users ( $p < 0.001$ ).
H3 (Norm Setting - Escalation): If the first reply to a toxic comment is toxic, the remaining conversation will be more toxic.
- Finding: Supported. There is a positive, statistically significant correlation between the toxicity of the immediate reply and the toxicity of the rest of the thread ( $p < 0.001$ ).

4. Qualitative Analysis Results

To understand the nature of "standing up" (intervening), the authors manually labeled a subset of 61 conversations where the first toxic reply targeted a human.

Positive Standing Up: Defined as respectful correction, setting boundaries, or defending the target.
Result: Positive standing up occurred in only 19.7% of cases (20 out of 1,440 replies).
Implication: Users rarely attempt to resolve conflicts; instead, they often ignore the toxicity or escalate it.

5. Significance and Implications

Social Norms as Cues: The study provides empirical evidence that online behavior is heavily influenced by perceived social norms. Silence or inaction by a group is interpreted as implicit approval of toxicity, creating a feedback loop that disinhibits further abuse.
The Critical Window: The immediate response to the first toxic comment is a pivotal moment. A non-toxic response can establish a norm of civility, while a toxic response triggers a cascade of further abuse.
Intervention Design: The findings suggest that platform interventions should focus on:
1. Reducing Bystander Diffusion: Encouraging early intervention by making users feel personally responsible rather than assuming "someone else will act."
2. Prompting Positive Norms: Designing systems that encourage or highlight non-toxic responses immediately after a toxic event to set a constructive tone for the rest of the thread.
Limitations: The study is limited to English tweets, relies on API-based toxicity detection (which has known biases), and uses participation counts as a proxy for observation (since true view counts are not unique).

Conclusion

The paper demonstrates that group dynamics play a decisive role in the evolution of toxic conversations on Twitter. The presence of many observers before a toxic event reduces the likelihood of intervention, and the tone of the first response acts as a powerful social cue that determines whether a conversation spirals into further toxicity or remains civil. Understanding these mechanisms is essential for developing effective, norm-based interventions to mitigate online harassment.