How MagenDetect works
A four-stage pipeline that distinguishes reporting on antisemitism from exhibiting it — the critical capability most keyword-based tools lack.
Dog-whistle scan
47 hardcoded patterns in English and Hebrew, each mapped to specific IHRA Working Definition examples. Runs instantly with no API call. Identifies coded antisemitic language like "globalist," "dual loyalty," and "zionist entity."
0ms · No external dependencies
Triage
GPT-4o-mini makes a fast relevance check: does this content mention Israel, Jewish topics, or related subjects? Irrelevant content (sports, weather, finance) short-circuits here — saving the cost of deep analysis.
~1.5s · GPT-4o-mini
Three-track analysis
GPT-4o performs deep analysis across three independent tracks: antisemitism (8 signal types), anti-Israel bias (7 signal types), and context omissions (6 signal types). Each piece of evidence gets an attribution label: publisher voice, quoted speech, rebuttal, or unclear.
~4s · GPT-4o
Adversarial verification
A skeptical reviewer challenges every signal from Stage 2. It asks: is this truly in the text, or is the analyst over-reading? Signals that are clearly unsupported are rejected. Ambiguous ones get downgraded confidence scores rather than blanket rejection.
~3s · GPT-4o
Calibration
Attribution-weighted scoring applies multipliers that reflect how problematic content really is based on who said it:
An article that quotes an antisemitic statement in order to condemn it scores 85% lower than an article whose own voice uses the same language. This single rule is what separates usable journalism tooling from a keyword grep.
Ready to integrate?
Get an API key and start scanning content in minutes.