Someone who encounters the word “antisemitism” in a headline and tries to understand what it means tends to picture explicit slurs, the most recognizable form of the language. But research finds the explicit slur is the smallest part of it: in one study of more than 1.26 million QAnon subreddit posts, explicit antisemitic terms appeared in only 0.66% of posts while coded, implicit ones appeared in 8.6%, roughly thirteen times as often (Weinberg and colleagues, 2025, peer-reviewed). Most antisemitism is written in code, substituted words, dogwhistles, and symbols built to carry one meaning to a general audience and another to an in-group, and to slip past the filters that catch slurs. This report defines the term, then reads the language itself, how the code works, how it evolves to evade detection, and how its volume has moved, because that is where the phenomenon is now measured.
Key Findings
- Coded antisemitism far outweighs the explicit slur: in QAnon communities explicit terms appeared in 0.66% of posts and implicit, coded ones in 8.6%, and 95.1% of those implicit terms invoked an antisemitism dimension (Weinberg and colleagues, 2025, peer-reviewed).
- The code is built to evade detection: harmful content carrying dogwhistles slipped past automated toxicity scoring in the first large-scale study of coded rhetoric, which catalogued over 300 dogwhistles (Mendelsohn and colleagues, 2023, peer-reviewed), and ordinary-looking code words such as the substitution in which “Skype” stood in for a Jew were engineered to dodge keyword filters (Taylor and colleagues, 2017, a preprint).
- The lexicon mutates to stay ahead of moderation, so researchers now use large language models to surface emerging coded terms rather than fixed word lists (Kikkisetti and colleagues, 2024, a preprint).
- It is measurably hard to read, even for trained coders: agreement on whether a word was being used hatefully was higher in context (Cohen’s kappa 0.871) than by keyword (0.676), evidence that recognizing coded antisemitism is contextual, not a fixed list (Taylor and colleagues, 2017).
- Attention to it has ratcheted up, not receded: US search interest in “antisemitism” stepped from a 2015 to 2017 baseline near 5 to a level around 36 by 2023 and stayed there, spiking at each flashpoint without returning to its old floor (Google Trends).
What is antisemitism?
It is hostility toward or prejudice against Jews as Jews, a prejudice old enough to predate the word for it. The most widely used standard for measuring it is the IHRA working definition, the International Holocaust Remembrance Alliance’s framework, which most of the content studies here apply to code whether a given piece of text is antisemitic. It is a widely adopted but itself contested standard, a point this report returns to at the end.
The harder question is not what antisemitism means but how you find it, because most of it no longer says “Jew.” It speaks in substitutes, symbols, and coded phrases. So the rest of this report treats antisemitism the way the research now does: as a language with a surface and a code, and reads it.
Most antisemitism is coded, not explicit
The explicit slur is the visible tip. In an analysis of more than 1.26 million posts across two QAnon subreddits, explicit antisemitic terms appeared in 0.66% of posts, while implicit, coded terms appeared in 8.6%, about thirteen times more often, and 95.1% of the implicit terms appeared in sentences that invoked an antisemitism dimension or conspiracy narrative (Weinberg and colleagues, 2025, peer-reviewed). Reading antisemitism means reading the 8.6%, not the 0.66%.
That coded layer is not random. It organizes into a handful of mechanisms, each a way of saying “Jew” without saying it: the conspiracy-of-hidden-power trope and its proper nouns, ordinary words repurposed to dodge filters, a punctuation symbol, political dogwhistles, and slurs that route through Israel. The decoder below maps the most common terms in each; click a mechanism to zoom in, and hover any term for what it encodes.
How the code evades detection, and how it evolves
The coding is not incidental, it is the point. In the first large-scale computational study of dogwhistles, coded expressions that carry a benign meaning to most readers and a hateful one to an in-group, harmful content containing them slipped past automated toxicity detection, and the authors catalogued a glossary of over 300 such terms, the antisemitic archetype being “cosmopolitan,” which reads as “worldly” to most and “Jewish” to a few (Mendelsohn and colleagues, 2023, peer-reviewed). The substitution codes work the same way: words like “Skype,” given an alternate meaning by an extremist community, keep an innocent surface so a keyword filter sees nothing (Taylor and colleagues, 2017, a preprint).
Because detection chases the code, the code keeps moving. One 2024 pipeline uses large language models to surface emerging coded antisemitic terms precisely because the lexicon evolves to stay ahead of keyword lists (Kikkisetti and colleagues, 2024, a preprint). The triple-parenthesis “echo” shows the drift in miniature: it began as a way to tag named individuals of Jewish heritage and broadened into a mark on abstract targets such as “bankers” and “globalists” (Arviv and colleagues, 2021, peer-reviewed). The meanings are close enough to measure: in one platform’s word-embedding space the term “jew” sat nearest to the slurs “kike” (0.81) and “yid” (0.62), the coded and the explicit clustering together (McIlroy-Young and Anderson, 2019, peer-reviewed).
Is the coded language spreading?
Attention to it has stepped up and stayed up. US search interest in “antisemitism” sat near a value of 5 from 2015 to 2017, rose through the Pittsburgh synagogue attack in 2018 and again around 2022, and reached its highest sustained level after October 2023, never falling back to its earlier floor (Google Trends). Each flashpoint spikes the line, but the baseline between spikes keeps ratcheting up.
Search interest measures attention, a different thing from the volume of antisemitism itself, so it matters that other measures point the same way. The frequency of antisemitic terms on fringe platforms rose sharply over time, the word “jew” climbing 16-fold and the slur “kike” 61-fold across one dataset (Zannettou and colleagues, 2020, peer-reviewed); reported antisemitic incidents in the United States rose 344% over five years to the highest annual count on record (Anti-Defamation League, 2025, a monitoring organization); and the elevated level after October 2023 did not recede the following year (Tel Aviv University Kantor Center, 2026). These are three different facets, attention, language frequency, and reported incidents, and the report on how antisemitism rises around specific conflict events treats the event spikes in detail; here the point is the convergence.
How researchers read it, and where the definition is contested
Reading coded antisemitism is contextual work, and it is measurably hard. When annotators judged whether an ordinary-looking word was being used hatefully, they agreed far better with the surrounding context than from a keyword alone, at a Cohen’s kappa of 0.871 in context against 0.676 by keyword (Taylor and colleagues, 2017). That gap is the whole problem in one number: the same word is antisemitic or innocent depending on who is speaking and to whom, which is why a fixed dictionary of slurs misses most of it and why the measurement relies on expert, context-aware coding against a stated framework, usually the IHRA definition.
That framework is itself contested, and the dispute is mostly about its Israel-related clauses, which count some expression about Israel as antisemitic and exempt ordinary criticism of it. A competing standard, the Jerusalem Declaration on Antisemitism, draws that line differently. This report does not adjudicate where the line falls; the measured relationship between antisemitism and attitudes toward Israel is taken up in a separate report on whether opposition to Zionism is antisemitism.
Methodology and limitations
The evidence here is of three kinds. Content-analysis studies code samples of online text for explicit and implicit antisemitism, by expert annotation or model, and report shares (Weinberg and colleagues, 2025; Jikeli and colleagues, 2022). Method and code-word studies catalogue the coded lexicon and test detection (Mendelsohn and colleagues, 2023; Taylor and colleagues, 2017; Kikkisetti and colleagues, 2024; Arviv and colleagues, 2021). And volume measures track antisemitism over time by term frequency, reported incidents, and search attention (Zannettou and colleagues, 2020; Anti-Defamation League, 2025; Google Trends).
Three limits bound the reading. Coded meaning is inferred, and inference is probabilistic; the decoder above is illustrative of well-documented terms, not an exhaustive or fixed list, because the lexicon changes. The trend measures capture different facets and none alone proves a secular rise, though they converge. And the antisemitic share depends on the platform and the coding framework, so a figure describes its sample and its standard as much as the phenomenon.
Conclusion
So when a reader goes looking for antisemitism, does the slur they expect to find turn out to be the whole of it? The slur is real and measurable, but it is the smallest part: most of the phenomenon is written in code, and reading it means reading the 8.6%, not the 0.66%.
The numbers draw a consistent portrait across every axis this report measured. The coded form outnumbers the explicit slur by more than thirteen to one, 8.6% of QAnon-subreddit posts against 0.66%, and 95.1% of those implicit terms appeared in sentences invoking an antisemitism dimension. The code is engineered to evade: harmful content carrying dogwhistles slipped past automated toxicity scoring across a glossary of over 300 such terms, and a swap like “Skype” for a Jew kept an innocent surface to dodge keyword filters. It is measurably hard to read, agreement running higher in context than by keyword, a Cohen’s kappa of 0.871 against 0.676. And attention has ratcheted up rather than receded, US search interest stepping from a baseline near 5 to around 36 by 2023 and never falling back, while term frequency on fringe platforms rose 16-fold for “jew” and 61-fold for “kike.”
That coded language does not stay in the dataset. It lives in the everyday talk the reader already recognizes: the aside about globalists secretly steering events, the meme casting Soros as a puppet-master, the triple parentheses dropped around a name, the worldly-sounding word that means one thing to most and another to a few. The Institute measures cultural and online portrayal because it is one of the inputs that shapes how a group is seen. What a count can fix in place is the 0.66% and the 8.6%, the glossary of over 300 terms, the “Skype” that once stood in for a Jew. What it cannot pin down is the next substitution, already drafted in some thread the dictionary has not reached, because the lexicon keeps mutating a step ahead of the tools built to read it. The measurement holds the words that have surfaced; the question that outruns it is whether a language engineered to say “Jew” without saying it, multiplying faster than anyone can catalog, is one of the currents that feeds antisemitism downstream. That is the open question the numbers leave standing for a society to weigh.
Frequently Asked Questions
What is the difference between explicit and coded antisemitism?
Explicit antisemitism uses recognizable slurs or open hostility; coded, or implicit, antisemitism encodes the same meaning in substituted words, symbols, or dogwhistles that look innocent to outsiders. In QAnon communities, the coded form was about thirteen times more common than the explicit form, 8.6% of posts against 0.66% (Weinberg and colleagues, 2025).
What is a dogwhistle?
A dogwhistle is a coded expression that carries one meaning to a broad audience and a second, hidden meaning to an in-group, used to evade both social consequences and content moderation. A study that catalogued over 300 of them gives the antisemitic archetype “cosmopolitan,” which reads as “worldly” to most and “Jewish” to a few (Mendelsohn and colleagues, 2023).
What is the (((echo)))?
The “echo” is a set of triple parentheses placed around a name to mark the person as Jewish. Researchers tracked how it spread from tagging named individuals to marking abstract targets such as “bankers” and “globalists” (Arviv and colleagues, 2021).
Is antisemitism actually increasing, or just the attention to it?
Search attention has clearly stepped up and stayed up (Google Trends), and that is attention, not the thing itself. But separate measures of the thing itself point the same way: antisemitic term frequency rose many-fold on fringe platforms (Zannettou and colleagues, 2020) and reported incidents rose 344% over five years (Anti-Defamation League, 2025). The measures converge without any single one being definitive.
What is the IHRA definition?
The IHRA working definition is the International Holocaust Remembrance Alliance’s framework for identifying antisemitism, the standard most of the content studies here use to code text. It is widely adopted and itself contested, mainly over its clauses about Israel-related expression, with a competing standard, the Jerusalem Declaration, drawing that line differently.
Sources
- Anti-Defamation League, 2025. Audit of Antisemitic Incidents 2024. adl.org. Monitoring organization.
- Arviv, Hanouna, Tsur, 2021. It’s a Thin Line Between Love and Hate: Using the Echo in Modeling Dynamics of Racist Online Communities. ICWSM 2021, arXiv:2012.01133. Peer-reviewed.
- Kikkisetti, Mustafa, Melillo, Corizzo, Boukouvalas, Gill, Japkowicz, 2024. Using LLMs to Discover Emerging Coded Antisemitic Hate-Speech in Extremist Social Media. arXiv:2401.10841. Preprint.
- McIlroy-Young, Anderson, 2019. From “Welcome New Gabbers” to the Pittsburgh Synagogue Shooting: The Evolution of Gab. ICWSM, arXiv:1912.11278. Peer-reviewed.
- Mendelsohn, Le Bras, Choi, Sap, 2023. From Dogwhistles to Bullhorns: Unveiling Coded Rhetoric with Language Models. ACL 2023, arXiv:2305.17174. Peer-reviewed.
- Jikeli, Axelrod, Fischer, Forouzesh, Jeong, Miehling, Soemer, 2022. Differences Between Antisemitic and Non-Antisemitic English Language Tweets. Computational and Mathematical Organization Theory, DOI 10.1007/s10588-022-09363-2. Peer-reviewed.
- Paudel, Blackburn, De Cristofaro, Zannettou, Stringhini, 2021. Soros, Child Sacrifices, and 5G: Understanding the Spread of Conspiracy Theories on Web Communities. arXiv:2111.02187. Preprint.
- Taylor, Peignon, Chen, 2017. Surfacing Contextual Hate Speech Words within Social Media. arXiv:1711.10093. Preprint.
- Tel Aviv University Kantor Center, 2026. Antisemitism Worldwide Report for 2025. Tel Aviv University. Academic monitoring center.
- Weinberg, Levy, Edwards, Kopstein, Frey, and colleagues, 2025. Hidden in Plain Sight: Antisemitic Content in QAnon Subreddits. PLoS ONE, DOI 10.1371/journal.pone.0318988. Peer-reviewed.
- Zannettou, Finkelstein, Bradlyn, Blackburn, 2020. A Quantitative Approach to Understanding Online Antisemitism. ICWSM 2020, arXiv:1809.01644. Peer-reviewed.
- Google Trends, 2026. Search interest for “antisemitism,” United States, 2015 to 2025. trends.google.com. Primary data source.