New York City Police Department (NYPD) vehicles are seen in Brooklyn, New York, United States, on Oct. 13, 2024. Photo: Kyle Mazza via Reuters Connect

Last Friday night, during Shabbat services at Central Synagogue in Manhattan, a man walked toward the front of the sanctuary, shouted antisemitic slurs, struck a congregant, and headbutted a security guard before being taken into custody. He faces multiple charges, including hate crimes. Three weeks earlier, on Tisha B’Av, a visibly Jewish man wearing a kippah, who had just left services, was stabbed with a screwdriver on the Upper West Side by an attacker shouting “Allahu akbar.” According to NYPD data, confirmed anti-Jewish hate crimes in New York City rose 53 percent this July compared to July of last year; through the end of July, Jews were the targets of 57 percent of all confirmed hate crimes in the city.

Each of these events will eventually be recorded as what it legally is: one incident. But an incident contains far more information than the count preserves. It has a victim and a perpetrator, a location and a weapon. It also has a discursive history — the narratives that circulated before it, the commentary that follows it, the justifications that outlive it. The count is not wrong to record one incident. It is measuring a different object than the one we increasingly need to observe.

The FBI’s new Hate Crime Statistics report, whose findings The Algemeiner reported this week, is the national ledger of such incidents. It found that Jews, 2 percent of the US population, were the targets of 63 percent of all religion-based hate crimes in 2025 — 1,528 incidents, down from 1,938 in 2024, and roughly seven times the number recorded against the next most targeted religious group. Other measurements complicate the apparent improvement. The ADL’s annual audit, which applies a broader definition covering harassment, vandalism, and assault, recorded 2025 as the third-highest year in its history and found a 39 percent increase in assaults involving a deadly weapon.

It was the first year since 2019 in which American Jews were murdered in antisemitic attacks and, according to the J7 task force, the deadliest year for diaspora Jews worldwide in more than three decades. ADL chief Jonathan Greenblatt warned in his response to the FBI data against letting “a modest yet positive decline in the topline domestic numbers” obscure the persistence of the threat. The warning is justified, and it can be given an empirical basis. These datasets measure different things: reported crimes meeting law-enforcement classification standards in one case, a wider universe of incidents in the other.

The FBI’s statistics can tell us how many incidents were reported and how they were classified. They cannot tell us how prevalent antisemitic expression is, which justificatory narratives are spreading, whether a conspiracy frame is gaining ground, or how quickly an attack generates new discourse. This is not a criticism of the FBI; its instrument was never designed for those questions. But a declining crime count cannot tell us whether the environment that produces the crimes is improving. To know that, we would need to measure the environment itself.

What precedes the counted incident

Before antisemitic violence becomes a counted incident, it has a communicative history — words, images, memes, symbols, and the combinations among them — and much of that history is visible long before the violence to anyone equipped to read it. The perpetrators of Pittsburgh, Christchurch, and Halle were immersed for years in digital environments dense with the narratives, slogans, and iconography that later appeared in their attacks. Elias Rodriguez, who federal authorities charged with murdering two people outside the Capital Jewish Museum in May 2025, published a manifesto hours beforehand that framed the killings as legitimate protest and called for further “armed demonstrations.”

The claim here is not that a particular comment radicalized a particular perpetrator. That is often impossible to establish at the level of an individual comment, and it is not the relevant question. The claim concerns discursive continuity: the justificatory grammar these perpetrators articulate — violence against Jews as deserved, inevitable, a form of resistance — circulates in mass discourse before they act. Whatever else radicalization involves, its ideological component has to be encountered, articulated, and reinforced somewhere.

Quantitative studies in Germany, the United Kingdom, and Spain have found that surges of online inflammatory language precede, and improve the prediction of, offline hate crimes; the researchers behind these studies are careful to note that prediction is not causation, and that caution is worth adopting. We do not need to prove that discourse causes the next attack to show that incident statistics do not capture a layer of the phenomenon that may matter for anticipating it.

This circulation runs in more than one direction. From the top down: in our reception studies, phrases about Jewish-associated institutions that carry a long antisemitic genealogy — hidden money, monstrous agency, a single malign purpose — were extended by audiences from the named institution to Jews as a group, while phrases from the same speeches without that genealogy produced no uptake at all, regardless of the speaker’s intent or political camp. And from the bottom up: comment sections sharpen ambiguous formulations into explicit hostility, users validate and intensify one another, and engagement-driven algorithms amplify whatever provokes, from whichever direction it arrives. What incident statistics sit atop is not a broadcast with a sender and receivers but an ecosystem in which elite speech, participatory dynamics, and platform architecture continuously feed one another.

Most of this layer is not a criminal offense at all, and much of it is not antisemitic on its surface. It may say “Zionists” or “globalists”; it may work through memes, emojis, and wordplay. None of these expressions is antisemitic in itself. The analytical question is when such terms operate within antisemitic constructions — attributions of hidden collective power, conspiratorial agency, essential malice — and answering that question requires context, not keywords. The Decoding Antisemitism project, which I led from 2020 to 2025, annotated more than 300,000 items of digital content and published a research lexicon documenting these constructions, downloaded nearly half a million times since 2024. Keyword-based detection systems are poorly suited to most of what that lexicon documents, because the individual comment retains plausible deniability while the cumulative message remains legible to its intended audience.

What follows it

The gap becomes concrete when one examines what a single counted incident generates. In March, a man rammed a truck into the entrance of Temple Israel in West Bloomfield, Michigan, one of the largest Reform synagogues in the country; synagogue security stopped the attack. In the FBI’s next report, it will appear as one incident. In one of our rapid-response analyses, my team examined 1,600 comments posted within 24 hours beneath the YouTube coverage of eight major US news outlets. In our coding, roughly one in six comments was antisemitic, and the most frequent antisemitic response category was neither celebration nor grief but the claim that Jews had staged the attack: “Mossad false flag operation.” “Fake to push the antisemitism narrative.” Alongside the staging claims ran the justifications — the attack described as expected, as provoked, as deserved.

We have applied the same protocol to five major attacks, from the Washington shooting through the attempt at the White House Correspondents’ Dinner in April, and the staging claim appears as a leading response category in each of these corpora. Across the attacks we have studied, a counted incident has repeatedly produced an uncounted wave of antisemitic communication in mainstream digital spaces, including the same justificatory narratives that appear, in the perpetrators’ own writings, before the violence. Whether and how strongly the aftermath of one attack conditions the next is a question our data cannot answer. What can be said is that there is no national reporting system designed to measure it.

One finding from this series illustrates why surface indicators can mislead. At the Correspondents’ Dinner shooting, explicit antisemitic content fell to near zero across most outlets we analyzed. The underlying conspiracy structure did not: claims of a staged event and a hidden orchestrator surged to record levels, with the identity of the supposed orchestrator reassigned — and at one channel whose audience already carried the older frames, claims of Mossad involvement returned immediately, without any pretext in the event itself. We observe these structures in left-leaning, centrist, and right-leaning comment ecosystems alike. A low reading on the surface indicator says little about the structure underneath, which is precisely the problem with measuring only what is explicit, or only what is criminal.

American Jews, meanwhile, respond to the environment rather than to the official numbers. In a Combat Antisemitism Movement survey released in July, 57 percent said they hide markers of their identity to avoid becoming targets — a comprehensible response in a year in which a visibly Jewish man was attacked on a Manhattan street. Notably, 32 percent said they refrain from posting content on social media that would identify them as Jewish: they are withdrawing from the very spaces in which the discourse described above circulates freely. The visibility problem now runs in opposite directions: antisemites can obscure what they are expressing, while Jews feel compelled to obscure what identifies them. Neither is captured by the incident count. They are not equivalent — one is the adaptation of those who spread hatred, the other the self-protection of its targets — but both mark the distance between the phenomenon and the instrument.

Measuring the environment

Responding to the FBI data, Shurat HaDin’s Nitsana Darshan-Leitner told The Algemeiner that officials must do “more than just tabulating statistics.” She is right, and the missing element has a concrete architecture. It cannot begin with automation, because one cannot scale what one has not first understood. It begins with granular, qualitative analysis of how antisemitism is actually expressed: the coded vocabulary, the conspiracy structures, the justification narratives, the combinations of image and text.

These patterns are then formalized into transparent taxonomies; the taxonomies guide expert annotation of large corpora; and the annotated corpora become the training and benchmarking ground on which computational models can be developed and, crucially, tested — so that we learn, rather than assume, whether the detection of implicit and coded hostility can work at scale. Built in this order, discourse-level measurement could accomplish what incident statistics cannot: track justificatory narratives before they crystallize into a manifesto, register the conspiracy surges that follow each attack, and indicate where the dynamic between the two may be accelerating.

This is the approach my own team is pursuing, and others should build competing versions of it — the field needs more of this work, not a monopoly on it. At NYU’s Center for the Study of Antisemitism, the Decoding Hate project, developed in close collaboration with AddressHate, is building this pipeline — from qualitative pattern analysis through taxonomy development and expert annotation to the systematic evaluation of detection models — across antisemitism and adjacent forms of hate in US digital discourse.

The question the FBI report raises is not whether the numbers went up or down. It is whether we are prepared to observe the layer in which the next numbers are being prepared.

Matthias J. Becker, PhD, is AddressHate Senior Research Scholar at NYU’s Center for the Study of Antisemitism, where he leads the Decoding Hate project. He led the Decoding Antisemitism project (2020–2025) and is Editor-in-Chief of the Digital Hate Review.