How the briefings are made
This is the evergreen pipeline behind every issue of Turning the Tide. Per-issue appendices note only what changed that month. The audience for this series checks methodology — so we keep it complete.
Posts classified as antisemitic are collected across the following platforms and forum ecosystems:
Data is collected using Brandwatch, Open Measures, the YouTube and Telegram APIs, and the Meta Content Library. Facebook and 4chan data are down-sampled to 20% due to export limits. Where engagement data is available, posts are collected in full; where it is not, posts are down-sampled.
Collection scope varies by account type. For forums such as 4chan, where comment volume is high and not all content comes from selected accounts, a targeted keyword filter is applied. For individual influencer accounts, a broader approach captures any content relevant to Jewish people or Judaism. On Telegram, YouTube and Facebook, comment-section data is also collected to support audience analysis.
ISD developed LLM classifiers to identify antisemitic content, grounded in the IHRA definition. Two models were evaluated: OpenAI’s GPT-4o-mini, used during influencer identification, and Anthropic’s Claude Haiku, deployed for the main analysis dataset after achieving stronger performance. Classifier performance was validated by manually coding a random sample of posts and comparing against model outputs.
Illustrative values
Using roughly 1,000 keywords drawn from existing ISD keyword lists spanning the ideological spectrum, ISD collected over 650,000 posts across 30+ platforms between September 2025 and February 2026. Posts with engagement were collected in full; where engagement data was unavailable, posts were down-sampled. The GPT-4o-mini classifier was then applied, producing an initial pool of tens of thousands of accounts that had posted antisemitic content.
From the initial pool, ISD focused on high-volume posters — a working list of over 2,000 accounts and forums — then applied a minimum-posting threshold: at least 10 flagged posts, reduced to 5 on platforms with thinner coverage such as Facebook. Some platforms, notably Gab and TikTok, remained underrepresented despite this.
To supplement the keyword-and-classifier approach, network-analysis techniques surfaced high-traction influencers it may have missed — examining the accounts most frequently engaged with by those already in the pool. The goal was to surface significant, influential accounts as neutrally as possible across the ideological spectrum, not to produce an exhaustive list.
Posts flagged as antisemitic by the Claude Haiku classifier — 305,000 in total — were imported into ISD’s proprietary data system, which normalises content across platforms into a common architecture, enabling cross-platform analysis in a unified environment.
ISD ran a clustering exercise using a topic-modelling approach in Nomic Atlas, producing a set of topics that analysts could review and characterise. This process produced 320 distinct clusters, relabelled into 12 broad narratives — allowing analysis of how those narratives pervade multiple online ecosystems at once.