Methodology

This tracker is based on a dataset of English-language replies from @grok on X in response to users tagging it with fact-checking requests. Data was collected using the approach described in Renault, Mosleh & Rand ("@Grok Is This True? LLM-Powered Fact-Checking on Social Media," 2026).

For each day, the tracker identifies the tweet that received the highest number of unique Grok fact-checking replies within 24 hours of the tweet's publication — the day's "main character." A tweet is assigned to the calendar day (UTC) on which it was posted.

Grok's verdict for each tweet is classified programmatically based on the language of its replies (e.g., "No, this claim is false" → false; "Yes, reports confirm" → true; partial confirmations or ambiguous responses → mixed). The classifier uses negation-aware pattern matching: it focuses on the opening sentence of each reply, checks for strong denial signals before affirmative ones, and treats any reply with both confirming and denying language as mixed. The overall verdict is the majority classification across all unique replies in the 24-hour window. The "Median Grok reply" shown on each card is the chronologically middle reply (50th percentile by timestamp), selected to represent the settled consensus rather than early hot-takes or late stragglers.

Community Notes data is sourced from X's public data dumps and may lag behind the Grok reply data by several days. Notes are matched to tweets by tweet ID and categorized by their current status (helpful, needs more ratings, or not helpful).