Replication materials for "The Enforcement and Feasibility of Hate Speech Moderation"

Code and derived data to reproduce every analysis, figure and table in Tonneau et al., "The Enforcement and Feasibility of Hate Speech Moderation". Running bash run_all.sh rebuilds all 36 supplementary tables, the cost and coverage frontiers and the figures in about two minutes on a laptop. The study audits hate speech enforcement on Twitter (now X) using a complete snapshot of public tweets posted in the 24 hours from 15:00 UTC on 20 September 2022, requeried five months later (February 2023) to measure tweet removal and account suspension. It also evaluates automated hate speech detection and simulates a human-AI moderation pipeline to estimate the cost of large-scale enforcement. Contents run_all.sh - one-command rebuild of the generated tables and figures code/ - analysis pipeline (sampling, annotation agreement, enforcement regressions, AI detection, human-AI simulation, figures) data/hateday_replication.parquet - 540,000 sampled tweets with human annotations, engagement metrics and the five-month enforcement outcomes data/scores/ - per-language classifier scores data/bootstrap_thought_experiment/ - human-AI simulation output data/external/ - tweet volumes, wages, moderator counts, label-only annotator votes, language counts of the full snapshot annotation/ - annotation guidelines results/ - generated tables and figures Raw tweet text is not redistributed, in line with the X/Twitter Terms of Service; records are keyed by tweet ID so text can be rehydrated by anyone with platform access. Account identifiers are replaced by salted hashes, and annotator identities are removed. Related resources. The underlying 24-hour collection is archived at GESIS (doi:10.7802/2516). The human annotations alone are released as the HateDay dataset on Hugging Face. The enforcement outcomes analysed in this paper are available only in this deposit.