Sky News investigated X's algorithm for political bias (news.sky.com)

Sky News ran a controlled audit of X’s recommendation algorithm and concluded it disproportionately amplifies right‑wing and extreme content — and content from politicians favoured by owner Elon Musk. To probe the platform’s opaque system (the public source code is partially redacted and researchers face steep API fees), the team created nine “clean” UK accounts (three left‑leaning, three right‑leaning, three neutral), simulated human engagement, and automated twice‑daily scrolling from 2–17 May 2025. That process produced 87,420 For You posts; Sky News reports the algorithm boosted posts with high engagement and video formats and fed users more right‑wing and extreme material. Technically, Sky News used a three‑phase workflow (account setup, metadata scraping, analysis) and scaled classification with Google’s Gemini 2.5 Flash LLM to label content type and author political alignment. They focused on the 5,981 authors whose posts comprised 67% of the dataset, applying an 80% confidence threshold and validating against human analysts. Agreement rates were high for political vs apolitical (97%/94%) and left vs right (93%/84%), but lower for “extreme” labels (82%/63%), highlighting classification uncertainty. For the AI/ML community the story underscores two points: LLMs can accelerate large‑scale platform audits, but reproducible, rigorous auditing is hampered by black‑box systems, redacted code, costly API access, and the inherent difficulty of reliably labeling edge cases like “extremism.”