Home
Settings

Dashboard Home

AddressHate · Discourse Intelligence

We measure hate online.

At AddressHate, we build tools to study and empower research related to hate. Our goals and work focus on bridging the infrastructure and knowledge gap between researchers, policy makers, and different disciplines. We aim to track and better understand how ideologies and radicalization transfer between online sources, social media, and AI chatbots. By tracking these sources, studying interactions on them, and building tools that exponentially speed this process up, we aim to better understand and prevent hate. This page showcases our preliminary analytics tools.

Monitored Platforms

Possible Antisemitic content over the last 30 days. These numbers are based on keywords, and not human or machine classified.

Full monitor

Trending Phrases

The terms and phrases with the most movement over the last week.

Published research

All studies
AI Evaluation · AddressHate

8,400

answers graded

How readily do frontier AI models repeat antisemitic misinformation?

18 models against 400 adversarial prompts on Holocaust denial, the conspiracy canon, and coded dog-whistles.

AddressHate · July 2026

AI Evaluation · AddressHate

10

models tested

How readily do frontier AI models repeat anti-Black racist misinformation?

10 models against 200 adversarial prompts on race pseudoscience and the distortion of slavery.

AddressHate · Run ab, July 2026

AI Evaluation · AddressHate

200

adversarial prompts

How readily do frontier AI models repeat anti-LGBTQ misinformation?

10 models against 200 adversarial prompts on “curing,” trans healthcare, and the “groomer” smear.

AddressHate · Run lgbtq, July 2026

Everything on this dashboard

Monitor

What is being said right now, and where.

Social Media

Cross-platform volume for our antisemitism query — X, Reddit, YouTube, Bluesky, Tumblr, the chan boards, and thousands of forums and news sites. Where the discourse concentrates, and which coded signals are gaining ground.

Open monitor

Evaluate

How platforms and AI models perform against it.

Platform Signal Grades

An A–F grade per platform for the density of antisemitic lexical signal in our query window. It measures observed signal risk relative to other platforms — it is not a platform-safety or moderation grade.

Compare platforms

AI Model SafetyPublished study

How well frontier language models detect antisemitic content, and how often they refuse to repeat antisemitic misinformation. Graded from our own adversarial evaluations, not vendor claims.

Providers graded: 10Models tested: 18Answers graded: 8,400Overall rejection: 86.2%
Compare models

Research

The methodology, the studies, and the people behind both.

ResearchPublished study

Our mission, our institutional partners, and every study we have published — each with its methodology and underlying data laid out in full.

Published studies: 4Partners: USC Shoah Foundation · NYU · Stanford IO
Read the research

TaxonomyReference

Every category we measure, defined in plain language with an example of how it shows up online. The antisemitism scheme follows the Decoding Antisemitism Lexicon.

Category groups: 6Concepts: 43
Browse taxonomy

FAQReference

Plain-language answers on where the data comes from, how content gets labeled, what the grades do and don't mean, and how to read the rest of the dashboard.

Read the FAQ

StaffReference

The researchers, engineers, and advisors behind the work, and how we work together.

Team: 18 people
Meet the team

Extension

The annotation tool that produces our ground truth.

For UsersReference

What the Annotator does, which sites it runs on, and how to get an account. It captures the reply chain, the quoted post, and the parent video alongside every label.

Supported sites: 13
See the extension

Why Annotate?Reference

Why hand-labeled context beats scraped text — how annotation feeds moderation tooling, and what gets lost when it is skipped.

Read the case