Bots and Cyborgs: Unmasking Digital Astroturfing in the 2017 Chilean Election

TL;DR

Social media is the new battleground for political influence, but not all soldiers are human. This study analyzes the 2017 Chilean Presidential Election, revealing how "cyborgs"—human-assisted automated accounts—were used to inflate candidate popularity. While machine learning models (SVM, Random Forest) showed promise in controlled training (83% accuracy), their real-world performance dropped to near 56%, highlighting the sophistication of modern political bots.

The "Cyborg" Menace: Beyond Simple Automation

The research identifies a critical evolution in social media manipulation: the Cyborg. Unlike traditional bots that are 100% autonomous, cyborgs are accounts controlled by humans using automation tools like TweetDeck. These accounts can coordinate massive "retweet storms" in milliseconds, granting a candidate false credibility while evading simple bot-detection filters that look for purely "robotic" timing.

The authors' intuition was triggered by a simple observation: during the September 2017 debates, certain candidates saw massive spikes in mentions (over 4,000 in an hour) that consisted almost entirely of retweets with zero original content.

Methodology: The 6-Dimensional Feature Space

To catch these digital ghosts, the researchers extracted 61 unique features categorized into six domains:

1. User-based: Profile metadata (followers, account age).
2. Friends: Patterns in mentions and retweeting behavior.
3. Network: The structure of hashtag and retweet interactions.
4. Temporal: The distribution of time intervals between tweets.
5. Content/Language: POS-tagging (verbs/adjectives) and entropy of the text.
6. Sentiment: The emotional intent (positive/negative) of the messages.

Figure 1: Comparison of candidate mentions during the debate shows abnormal spikes suggesting coordinated automation.

Why the Models Stumbled: The Gap Between Theory and Reality

The researchers tested four major algorithms: Random Forest, AdaBoost, Decision Trees, and Support Vector Machines (SVM).

The Performance Paradox

In the training phase (using a mix of Botometer-labeled and manually tagged data), the results were stellar:

• Accuracy: ~0.83 across all models.
• Precision/Recall: Highly balanced.

However, when applied to the actual Chilean testing set, the performance plummeted:

• SVM Accuracy: 0.56
• Random Forest Accuracy: 0.55

Table 1: The testing results showcase a significant "reality gap" in bot detection.

The Forensic Evidence

The data revealed that candidate Marco Enríquez-Ominami had a retweet-to-original-tweet ratio of roughly 6:1, significantly higher than his peers. Furthermore, a high percentage of his support came via TweetDeck, a platform that facilitates multi-account synchronization.

Critical Insight & Conclusion

The study proves that political actors are no longer using "dumb" bots. Instead, they utilize Hybrid/Cyborg systems that carry the "Inductive Bias" of human behavior but the "Scale" of automation.

The failure of standard ML models to maintain high accuracy during testing suggests that:

1. Context Matters: Bot behaviors in a Chilean political context differ from the general "Global" bot datasets.
2. Feature Evolution: Static features (like followers/friends) are being "gamed" by bot creators. Future research must look deeper into Graph-based representations to map the hidden connections between these accounts.

Takeaway: As we approach more global elections, the ability to distinguish between a "volunteer" and a "coordinated cyborg" remains one of the hardest—and most vital—challenges in computational social science.