Navigating the intricate landscape of online content moderation, Twitter is known to be one of the largest platforms when it comes to discussions of any topic. Twitter embraces the mission to “promote and protect the public conversation” as a platform for exercising people’s freedom of speech. However, given the many controversies on its platform, maintaining free speech on a huge platform like Twitter/X is a difficult task since it deals with protecting everyone's right and at the same time creating an environment to share thoughts. These challenges include misinformation, the evolution of moderation on Twitter to X, the protection of freedom of speech, a case of 2022 Lima elections showing hate of speech, and how we can develop a solution for flagging hate speech through trained models in order to make Twitter a safe place for freedom of speech.

🧠 Just wrapped up a study exploring whether Twitter/X actually creates echo chambers—and the results might surprise you. We tracked 243 users over 3 weeks, switching them between the algorithmic “For You” timeline and the chronological “Following” feed. Using a custom browser extension, we collected every tweet they saw and interacted with. Findings? The “For You” feed showed fewer news tweets, but they were more balanced, less extreme, and slightly more reliable than the “Following” timeline. In fact, the chronological feed exposed users to more ideologically extreme content—because people naturally follow like-minded accounts. What’s more, even when users saw more reliable content, their perceptions didn’t really change. Conclusion: It’s not always the algorithm creating echo chambers—sometimes, it’s just human nature. 🧵👇

research paper: Lower Quantity, Higher Quality: Auditing News Content and User Perceptions on Twitter/X Algorithmic versus Chronological Timelines. Proc.

🧵 Thread: Our recent study on Community Notes shows a critical flaw in misinformation management. While these notes help reduce engagement with misleading content, their delayed appearance is a major issue. We found that notes can take days to show up—whereas tweets typically go viral in just ~80 minutes. For instance, imagine a verified “doctor” tweeting false vaccine side effects. By the time a note is added, the damage is done. Our work suggests that rapid content spread combined with deeply entrenched beliefs means that slow fact-checking can inadvertently allow harmful misinformation to spread. More agile measures are needed.

#Misinformation #CommunityNotes

research paper: Did the Roll-Out of Community Notes Reduce Engagement With Misinformation on X/Twitter?

In our recent study, “Modulating Moderation – A History of Objectionability in Twitter’s Moderation Practices (2006–2022)”, we explored how Twitter evolved from a free speech-focused platform to a more politically responsive model of content governance. Initially, Twitter embraced a libertarian, non-interventionist stance aligned with First Amendment values. However, as the platform became central to major events like the 2016 U.S. elections, Charlottesville, and the COVID-19 pandemic, it faced growing pressure to address hate speech, disinformation, and extremism.

This led to the development of what we call a “modulated moderation” system—an adaptive approach based on real-time social and political threats. We identified three categories of objectionable content: constant (e.g., child sexual abuse, terrorism) which were strictly banned; variable (e.g., hate speech, abusive behavior) which could result in suspensions or deletions; and contextual content, which was addressed using softer tools like strikes, content demotion, or labeling. Notably, some flagged tweets could later be "redeemed," showing that moderation was often temporary.

Rather than abandoning free speech, Twitter restructured it into a crisis-responsive model. This modulation gave the platform a dual role: both as a facilitator of speech and a controller—shaping public discourse in a complex digital ecosystem.

research paper: Modulating moderation: a history of objectionability in Twitter moderation practices

In our latest study, “We found no violation!”: Twitter’s Violent Threats Policy and Toxicity in Online Discourse, we examine how Twitter’s narrow moderation practices can unintentionally silence marginalized voices. Although the platform promotes free speech, it often fails to address veiled or euphemistic threats that foster harm.

We analyzed three tweets targeting U.S. Congresswomen Ilhan Omar and Alexandria Ocasio-Cortez. The tweets used phrases like “meet your maker” and implied violence or sexual aggression. Despite their threatening nature, all were initially allowed to remain on the platform. This reveals Twitter’s moderation policy focuses too much on explicit language while overlooking context and real-world impact.

A striking example is Rechelle Ritchie’s report of a threat from Cesar Sayoc. Twitter dismissed it as non-threatening, yet Sayoc was later arrested for mailing bombs. This case highlights the dangers of relying solely on explicit intent in moderation.

We argue that Twitter must evolve its policy to consider not just keywords, but also social, cultural, and political context, as well as how messages are perceived by their targets. Without this shift, moderation policies risk enabling toxic discourse and undermining true freedom of expression.

research paper: “We found no violation!”: Twitter's Violent Threats Policy and Toxicity in Online Discourse

🧵 Our recent study on the 2022 Lima municipal elections reveals how Twitter often deepens political division rather than encouraging democratic discourse in Peru.

We analyzed 346,000 tweets from political candidates, media outlets, and highly active users. What we found was striking: discourse was largely negative, emotionally charged, and rarely deliberative.

Even traditional media—despite generating the most engagement—often triggered responses filled with insults, mockery, and attacks. Rational debate was overshadowed by disillusionment and frustration with the democratic process.

We also observed “othering” language, where users labeled left-leaning groups as “caviares,” “progres,” or “reds.” These polarizing narratives drowned out more reasoned or evidence-based arguments. In short, Twitter became a space not for dialogue, but for political cynicism and populist expression. A cautionary look at the role of digital platforms during elections.📱

📄 Full study: Discursive Practices during an Electoral Cycle: Public Opinion and Political Disillusionment on Twitter

#Peru #Elecciones2022 #TwitterPolitics #PoliticalDiscourse

🧵 At Mapúa University, our team explored how AI can improve content moderation on platforms like Twitter/X. The current systems often miss harmful content—or worse, silence the wrong voices. We introduced an optimized approach using BERT, a transformer-based language model, to go beyond basic keyword filters and better understand context, intent, and disguised toxic language. We tested our model on datasets like AHS and JUB. With class weighting and data cleansing, we saw major improvements in detecting toxic tweets—better precision and F1 scores. But challenges remain. In one case, the model missed nearly 2,000 toxic tweets. In another, it flagged too many non-toxic ones, raising over-censorship concerns. There's a trade-off between catching all toxicity and preserving open conversation—but this work is a big step forward. Smarter, context-aware models are key.

📄 Full study: Optimizing Language Model-Based Algorithms for Enhanced Content Moderation and Toxicity Classification Using Twitter Data

#AI #ContentModeration #NLP #TwitterData #MapuaUniversity

In conclusion, while misinformation often stems from users' resistance to changing their beliefs, tools like Community Notes can help though they need improvements to alert users before they engage with misleading posts. Over time, Twitter has shifted from a free speech model to a modulated moderation system, balancing content control with public and business pressures. However, incidents like the Cesar Sayoc case and the 2022 Lima elections highlight gaps in this system, especially in how harmful language is handled. To address this, researchers like Tomas, Carlos, Ebora, and Garin proposed a BERT-based model to detect disguised or coded toxic speech. While not perfect, it outperformed traditional filters and marks an important first step in building smarter, more effective content moderation tools.