AI in the Gaza War | German algorithms mistake criticism of Israel for anti-Semitism

Six years ago, the Center for Research on Antisemitism at the Technical University of Berlin launched the "Decoding Antisemitism" project. Its main goal was to develop an AI language model that detects antisemitic content in social media comment sections and reports it to the platform provider for deletion. According to the project leader, the project is not primarily concerned with antisemitism on the political right—its most widespread and dangerous form. The focus is on antisemitism among the middle class, which is expressed primarily in implicit, verbal forms.

To train the language model, "Decoding Antisemitism" collected 103,000 online comments, approximately two-thirds of which refer to Palestine and Israel. However, as is common practice in this field, the project does not make its data collection and analysis available to other researchers for critical examination of the methodological foundation and implementation. The few publicly available details reinforce the impression that "Decoding Antisemitism" is primarily concerned with sanctioning pro-Palestinian voices.

This is evident in the collection of 21,000 comments after October 7, 2023. According to "Decoding Antisemitism," around 2,400 (12 percent) of them were antisemitic: for example, because they accuse Israel of fascism, apartheid, and colonialism, or question its right to exist. Furthermore, it is antisemitic to refer to the "Boycott, Divestment, and Sanctions" (BDS) campaign—which, since its inception 20 years ago, has primarily targeted the territories occupied by Israel in violation of international law. Furthermore, the state should not be labeled racist or accused of genocide. Finally, it is antisemitic to blame Israel solely for the current suffering of the Palestinians in the Gaza war.

However, none of these claims and accusations are antisemitic in nature per se. Therefore, "Decoding Antisemitism" is of little use in drawing conclusions about the antisemitic motivations of the authors of the comments. One example of this questionable approach is the handling of comments about the killing of Palestinian children by the Israeli army. According to "Decoding Antisemitism," such accusations are antisemitic because they allude to the medieval Massacre of the Innocents—an antisemitic rumor originating in the 12th century that claimed Jews were kidnapping, torturing, and killing Christian children to use their blood for Passover.

How the killing of Palestinian children is supposedly connected to this Christian ritual murder legend is explained by the staff of "Decoding Antisemitism" in a "Guide to Identifying Antisemitism Online," a five-hundred-page glossary . However, the glossary denies the fact that the Israeli army has demonstrably been killing or seriously injuring Palestinian children and young people for decades—by the thousands currently in the Gaza War.

"Decoding Antisemitism" cites the "emotional intensity" of such comments as evidence of an antisemitic connotation—other possible motivations for why someone might react emotionally to the murder of children are not considered. Of course, it's conceivable that the authors were thinking about the child murder legend—but their comments alone are not proof of this.

A more serious and scientifically robust approach would examine the context of these comments more closely to determine their motivation. Without such an analysis, the conclusions would lead to the imputation of antisemitism to the authors in order to identify it in the next step—a circular argument. The AI developed by "Decoding Antisemitism" is primarily interested in phrases and buzzwords and is therefore not a scientifically sound methodology. This also makes it unsuitable for seriously combating antisemitism.