Research · Mission
AddressHate is dedicated to combating antisemitism and other forms of hate on social media by harnessing research and technology to identify sources of hate, understand the mechanisms that drive its spread, and develop effective tools to counter and prevent it. By combining rigorous social science with machine learning, we build open infrastructure that turns the flood of online content into clear, actionable evidence — surfacing emerging narratives, measuring how harmful speech moves across platforms, and giving researchers, platforms, and policymakers the tools they need to respond before hate takes hold.
Studies
Original research from our team and collaborators — each with its methodology and underlying data laid out in full.
Eighteen leading language models tested against 400 adversarial prompts on Holocaust denial and distortion, the classic and modern antisemitic conspiracy canon, coded dog-whistles, and tropes disguised as criticism of Israel — including the contested boundary with legitimate political criticism. Graded 0–4 by a three-model judge panel, with all 8,400 answers and 20,511 judge scores available to download.
AddressHate · July 2026
Ten leading language models tested against 200 adversarial prompts on the pathologization and “curing” of LGBTQ people, distortions of transgender healthcare, and the “groomer,” “social contagion,” and “gender ideology” smears — graded 0–4 by an AI judge panel, with every question, answer, and score available to download.
AddressHate · Run lgbtq, July 2026
Ten leading language models tested against 200 adversarial prompts on race pseudoscience, the distortion of slavery and Jim Crow, and racist conspiracies — graded 0–4 by an AI judge panel, with every question, answer, and score available to download.
AddressHate · Run ab, July 2026
A close reading of 1,000 YouTube comments across ten videos, examining how a host’s framing shapes the permission structure for antisemitic responses to an explicitly Nazi-referential song.
Matthias J. Becker, Isabelle Deutsch & Jameson Verser
Methodology
How we define, collect, and label the content behind every number on this dashboard.
Approach
The methods used to track online hate fall into a few broad families. Here is how ours differs at each stage — from how content is found, to how it's labeled, to how you can check the results.
| Manual reporting studies | Keyword + human vetting | AddressHate | |
|---|---|---|---|
| How content is found | Hand-picked by researchers | Keyword / dictionary matches | Engineered queries + coded-language detection, across fringe and mainstream platforms |
| How it's labeled | Grouped loosely as “hate” | A fixed list of definitional examples | A two-axis research taxonomy (ideation × category, ~46 tropes) that rolls up into any framework |
| How the model works | No model — manual review | A rented, general-purpose model or none | A classifier we train on our own expert annotations and can measure and improve |
| How much it sees | Hundreds of posts per study | Thousands of vetted posts over years | Hundreds of thousands of posts, growing continuously |
| How often it runs | A one-time snapshot | Ongoing, but limited by analyst time | Continuous automated monitoring, human-validated |
| Unit of analysis | The single post | The single post | The narrative — which outlasts single takedowns and bans |
| How you can check it | Not published | Accuracy rarely published | Accuracy and expert-agreement metrics published on the dashboard |
Contact
Questions about our data, methodology, or partnership opportunities? We'd love to hear from you.