Problem Statement: Prevalence of Spam Accounts on Twitter

Twitter defines four types of spam behavior in their policy: commercially-motivated spam, inauthentic engagements, coordinated activity, and coordinated harmful activity. These violations can be enacted by promotional accounts, false/fake accounts, bots, malicious accounts (trolls), and compromised accounts.

This report examines the prevalence of spam accounts on Twitter, evaluates Twitter's spam detection processes, and analyzes the effects of spam accounts on the platform. We will use multiple methodologies to estimate the true rate of spam accounts and assess whether Twitter's reported figures are accurate.

Methodology for Estimating Spam Account Prevalence

1

Count and Extrapolate

Sample mDAU accounts, identify spam, and extrapolate to full list

2

mDAU Modeling

Create machine learning model to distinguish spam from legitimate users

3

Twitter's Other Spam Systems

Apply Twitter's Smyte and Botmaker workflows to mDAU data

4

Scientific Consensus

Use state-of-the-art tools like Botometer to establish baseline

Key Findings on Spam Account Prevalence

1

Manual Review

Re-estimated mDAU spam account rate of 10.32% (vs Twitter's 0.72%)

2

Machine Learning Model

8.01% of active mDAU accounts predicted to be spam

3

Scientific Literature

Studies consistently report 9-21% spam accounts

4

Spam Activity

Spam accounts nearly twice as active as non-spam accounts

Issues with Twitter's Spam Evaluation Process

1

Uncertainty Handling

Process collapses uncertainty toward "Good" classification

2

Limited Information

Evaluators lack access to crucial non-public account data

3

Coordination Detection

Insufficient investigation of coordinated inauthentic behavior

4

Off-Platform Activity

Overlooks common cross-platform amplification tactics

Effects of Spam Accounts on Twitter

1

Scams and Fraud

Cryptocurrency schemes, counterfeit products, phishing

2

Stock Manipulation

Coordinated efforts to influence market perceptions

3

Political Disinformation

Distortion of discourse, amplification of false narratives

4

Public Health Misinformation

Spread of unverified medical claims, conspiracy theories

Recommendations for Improving Spam Detection

1

Update Training Materials

Regularly revise to address evolving spam tactics

2

Enhance Coordination Detection

Implement tools to identify inauthentic networks

3

Improve Annotator Assessment

Regularly evaluate and retrain spam detection staff

4

Cross-Platform Analysis

Consider off-Twitter activity in spam determination

Implications for Twitter's Business Model

The higher prevalence of spam accounts has significant implications for Twitter's advertising-based business model. Spam accounts generate a disproportionate amount of activity, potentially inflating engagement metrics presented to advertisers.

Additionally, the presence of spam accounts degrades the user experience for legitimate users, potentially impacting user retention and growth. Twitter must balance aggressive spam removal with the risk of falsely flagging legitimate accounts.

Conclusions and Next Steps

Our analysis suggests that Twitter's reported spam account rate significantly underestimates the true prevalence of spam on the platform. The company's spam evaluation process has several flaws that lead to systematic undercounting of spam accounts.

To address these issues, Twitter should:

1

Revise Methodology

Implement more robust spam detection techniques

2

Increase Transparency

Provide more detailed reporting on spam account prevalence

3

Invest in Technology

Develop advanced AI/ML tools for spam detection