Beyond the for You Page: Uncovering Algorithmic Bias in New York's Mayoral Race (medium.com)

Researchers analyzing TikTok’s recommendations say the platform’s recommender may be tilting the playing field in New York City’s mayoral race. Using reverse engineering (including a leaked onboarding doc and public algorithm disclosures), the team collected what real and synthetic user accounts were actually shown rather than scraping all posted content. They built a regression model that predicts a video’s views from metadata (R² = 0.928 for videos >1,000 views) to establish expected organic reach, then trained a classifier — “Excessive Publicity” — to label videos with unexplained over- or under-performance. Validation against paid ads found the classifier flagged 76.4% of labelled promotions, and the analysis focuses on group-level behavior (keyword clusters, narrative labels) rather than individual outliers. Key findings: across the dataset 17% of non-ad videos exhibited “Excessive Publicity,” but political content was a major outlier at 55%. Within political keywords, content tied to Zohran Mamdani showed materially higher rates of non-organic amplification (both pro‑Mamdani and anti‑Cuomo videos), while pro‑Cuomo content fell below the political baseline — consistent with relative amplification of one candidate and suppression of the other. The authors stress these are early, statistical results, but they underscore how opaque recommendation systems can shape public perception and voter exposure, highlighting the need for transparency, independent monitoring, and further scrutiny of algorithmic effects on democratic processes.