Angry but Accurate: Detecting and Profiling the Counter-Misinformation Ecosystem on Twitter

Researchers have long operated under a convenient assumption. If a social media post drips with intense, negative emotion, it is likely misinformation. This heuristic—treating emotional volatility as a proxy for falsehood—shapes how platforms moderate content. It also influences how researchers identify bad actors. However, a new study from the University of Southern California suggests this assumption is fundamentally flawed.

The researchers studied how people on Twitter push back against false COVID-19 claims. They found that people correcting misinformation actually use more negative emotions. Specifically, they use more anger and sadness than the people spreading the lies. These "correctors" also tend to be more established users. They often have more followers and older accounts. By analyzing a massive corpus of tweets, the study reveals that the "counter-misinformation ecosystem" is a highly charged environment of indignation.

Beyond the "False Equals Emotional" Heuristic

Current approaches to identifying misinformation often rely on detecting "outrage" or "fear" as signals of deception. While false content is frequently emotionally charged to encourage sharing, the existing literature has largely treated emotional intensity as a unique signature of falsehood. This creates a significant blind spot. It fails to account for the people who actively contest those falsehoods.

If a platform's moderation logic assumes that "high anger equals misinformation," it risks a massive collateral error. It may inadvertently silence the very users who work to correct the record. The authors note that while earlier studies hinted at this, they were often limited to small samples. They also relied on manual coding. There has been a lack of large-scale evidence showing whether the "angry corrector" is a consistent pattern across diverse claims.

A Pipeline for Stance and Sentiment

To map this ecosystem, the authors developed a multi-stage analytical pipeline .

The process begins by retrieving candidate tweets for 15,374 unique, fact-checked COVID-19 claims. They used BM25—a probabilistic search algorithm that ranks documents based on term frequency—to pull the 100 most relevant tweets for each claim.

The core of the detection engine is a Natural Language Inference (NLI) model. NLI is a task where a model determines the relationship between a "premise" (the fact-checked claim) and a "hypothesis" (the tweet). The model classifies the pair into one of three categories: entailment (the tweet supports the claim), contradiction (the tweet opposes the claim), or neutral. To train this model, the authors used Controllable Misinformation Generation (CMG). They prompted an LLM (specifically GPT-4o) to create realistic synthetic tweets that matched specific stances. This expanded their training data beyond what was manually labeled.

Once the tweets were classified, the authors moved to profiling. They compared 23 different features across the pro- and anti-misinformation groups. This included measuring discrete emotions using the Demux model, assessing toxicity with DeToxify, and calculating bot-likelihood via the Botometer X API.

The Signature of Indignation

The results challenge the status quo regarding emotional tone. The authors report that posts opposing misinformation are significantly more emotionally charged than those spreading it. Looking at the distributions of Ekman’s six basic emotions, the study finds that anti-misinformation tweets score higher in sadness, anger, and disgust .

While the authors characterize these differences as modest in magnitude, they are remarkably consistent. Using Cliff’s $\delta$—a statistical measure of effect size used to compare two groups—the paper reports the largest effect in sadness ($\delta = -0.125$). This is followed by anger ($\delta = -0.098$) and disgust ($\delta = -0.097$). The negative sign indicates these emotions are more prevalent in the "oppose" group.

The study also identifies a profile for the users driving this correction. The counter-misinformation ecosystem is anchored by more established accounts. These users tend to have more followers, appear on more Twitter lists, and possess older account ages.

Conversely, pro-misinformation tweets were found to be longer on average and contained higher levels of "surprise" .

Interestingly, bot scores did not meaningfully differentiate the two sides. This suggests automation is not uniquely concentrated in either camp.

Limits of the Signal

Several caveats remain. First, the findings are strictly bound to the COVID-19 era on Twitter. The heightened anxiety of 2020–2021 may have amplified these emotional signatures. These patterns might not generalize to other topics like climate change.

Second, the predictive power of these features is limited. The authors trained machine learning models, such as Random Forests, to see if they could distinguish a "corrector" from a "spreader." While the Random Forest achieved an F1 score of 0.632, the authors note this is a weak separation. This means the trends are real at a population level. However, you cannot reliably look at a single tweet and guess its stance based solely on its length or anger.

Finally, the reliance on pretrained models for emotion and toxicity detection introduces potential bias. Any errors made by the Demux or DeToxify models are inherited by the study. Furthermore, the authors admit that LLM-based data augmentation faces challenges. One issue is "task flipping," where the LLM accidentally generates a tweet that contradicts the intended label.

The Verdict: Don't Moderate the Anger

Is the "angry corrector" a reliable persona for engineers building moderation tools? Yes, but only if you stop using emotion as a shortcut for truth.

The study provides a clear warning for platform governance. If you build automated systems that downrank content based on high negative sentiment, you will likely suppress the voices of people fighting misinformation. The "angry but accurate" user is a fundamental part of the organic defense mechanism on social media. For developers, the takeaway is that emotional tone is a poor proxy for veracity. Effective moderation requires moving beyond simple sentiment analysis. It requires a deeper semantic understanding of whether a user is spreading a lie or expressing righteous indignation.

Figures from the paper