As part of Twitter’s efforts to strengthen platform integrity, the Curation team supported large scale spam labeling initiatives designed to improve the accuracy of automated detection systems. The project focused on identifying harmful or misleading content patterns that algorithms may struggle to detect without human context. Curators reviewed and labeled content across several categories including spam, Not Safe For Work content, coordinated promotions, and unhealthy engagement tactics. These labels were used to train and refine the platform’s moderation and recommendation systems so that harmful or low quality content could be detected more effectively at scale.
A key component of the work involved ensuring that definitions used by the algorithm reflected real world cultural and regional contexts. Because spam and promotional behaviors often manifest differently across markets, human reviewers helped surface where existing definitions were incomplete or misaligned with local content patterns.
What I did:
Content labeling and threat identification: I supported the spam labeling workflow by reviewing large volumes of content and accurately labeling posts against existing algorithmic guidelines. This included identifying spam signals, NSFW material, and promotional behavior designed to artificially boost engagement or manipulate platform visibility.
Regional context and policy refinement: Working within a global review framework, I flagged definitions that did not align with the content patterns commonly seen in my region (Sub-Saharan Africa). By surfacing these inconsistencies, I contributed to refining labeling guidance so that the algorithm could better interpret regional behaviors and reduce false positives or misclassification.
Improving algorithm training data: Accurate human labeling is critical for machine learning systems. My work contributed to building higher quality training datasets used to improve automated detection of spam and unhealthy platform behaviors.
Maintaining labeling consistency and quality: The task required careful adherence to review standards to ensure that labeling decisions remained consistent across reviewers and regions. This consistency helped maintain the reliability of the dataset used to train moderation models.
Outcome: This work strengthened the platform’s ability to detect and reduce spam, NSFW material, and manipulative promotional content across diverse global markets. By combining algorithmic review with human contextual judgment, the project helped improve how the system interprets regional content behaviors while supporting safer and healthier platform interactions.