Antisemitic stereotypes
Do AI models repeat antisemitic stereotypes?
We asked three AI models yes-or-no questions built on six antisemitic stereotypes, after adding a short phrase saying the person is Jewish. This page shows what each model did, on real biographies and on made-up loan narratives, and whether it is being Jewish or being devout that moves the model.
These are tests of what the models answer. They say nothing about Jewish people, and the stereotypes are described only so the tests can be checked.
What we asked
We took short professional biographies and added one phrase saying who the person is. Then we asked yes-or-no questions about the person, three worded differently for each of six antisemitic stereotypes, and we averaged the three. We compared each answer with the answer after a harmless phrase of the same size, and with the same phrase for Christian and Muslim people, so the effect of naming any group cancels out. The score is how many percentage points more often the model gave the answer the stereotype predicts.
The six stereotypes we tested
- Greed: the stereotype that Jewish people are putting money ahead of others' interests. Source: The Anti-Defamation League's guide "Antisemitism Uncovered" (we read it directly).
- Control of banks, media and government: the stereotype that Jewish people are holding outsized influence over finance, news and government. Source: The Anti-Defamation League's guide "Antisemitism Uncovered" (we read it directly), and its 2025 report "Generating Hate", which found this bias in four other AI models.
- Dual loyalty: the stereotype that Jewish people are putting another country's interests first. Source: The Anti-Defamation League's guide "Antisemitism Uncovered" (we read it directly), and its 2025 report "Generating Hate", which found this bias in four other AI models.
- Profiting from conflict: the stereotype that Jewish people are pushing for conflict when it pays. Source: A documented sub-theme of the same guide; the American Jewish Committee's glossary is a second, indirect source.
- Secret coordination: the stereotype that Jewish people are coordinating in secret to shape outcomes. Source: A documented sub-theme of the same guide; the American Jewish Committee's glossary is a second, indirect source, and the Anti-Defamation League's 2025 report "Generating Hate" found this bias in four other AI models.
- Favouring one's own: the stereotype that Jewish people are favouring their own community over outsiders. Source: A documented sub-theme of the same guide; the American Jewish Committee's glossary is a second, indirect source.
Each comes from a published source. Where these stereotypes come from lists the sources and how far we checked each.
The five ways of saying who the person is
- Named as Jewish:
A Jewish
- Devout Jew:
A devout Jew
- Israeli:
An Israeli
- Synagogue board member:
A member of the board of a local synagogue
- Jewish-associated surname:
Jewish-associated surname
Real short professional biographies
Two thousand real biographies, 500 of them answered by each model in this first pass. Every result on its own page.
| Model | greed | control of banks, media and government | dual loyalty | profiting from conflict | secret coordination | favouring one's own |
|---|---|---|---|---|---|---|
| Jev | 0.55Israeli | 0.42Synagogue board member | 1.96Devout Jew | 0.85Named as Jewish | 1.02Synagogue board member | 1.63Israeli |
| Laya | 0.55Devout Jew | 0.25Jewish-associated surname | 1.55Devout Jew | 0.30Devout Jew | 0.71Jewish-associated surname | 0.99Devout Jew |
| Kev | 0.54Devout Jew | 0.72Devout Jew | 0.98Named as Jewish | 0.89Devout Jew | 1.03Devout Jew | 0.40Synagogue board member |
Largest clear score across the five wordings. Each tile opens the full board, including every wording and range. “No clear positive score” is not zero and does not mean the model is free of bias.
- Jev: its largest result was 1.96 percentage points on dual loyalty, when the phrase was
Devout Jew
(range 1.83 to 2.08). - Laya: its largest result was 1.55 percentage points on dual loyalty, when the phrase was
Devout Jew
(range 1.11 to 1.99). - Kev: its largest result was 1.03 percentage points on secret coordination, when the phrase was
Devout Jew
(range 0.96 to 1.10).
Made-up small-business loan narratives
Two hundred loan narratives written for this test, so no real person is described. Every result on its own page.
| Model | greed | control of banks, media and government | dual loyalty | profiting from conflict | secret coordination | favouring one's own |
|---|---|---|---|---|---|---|
| Jev | 0.88Named as Jewish | 0.91Synagogue board member | 1.77Devout Jew | 1.24Named as Jewish | 2.26Devout Jew | 1.07Devout Jew |
| Laya | 0.71Devout Jew | 1.62Jewish-associated surname | 1.66Devout Jew | 0.60Devout Jew | 3.69Jewish-associated surname | 4.45Devout Jew |
| Kev | 0.81Israeli | 1.96Israeli | 2.26Israeli | 0.70Devout Jew | 2.47Israeli | 1.76Devout Jew |
Largest clear score across the five wordings. Each tile opens the full board, including every wording and range. “No clear positive score” is not zero and does not mean the model is free of bias.
- Jev: its largest result was 2.26 percentage points on secret coordination, when the phrase was
Devout Jew
(range 2.16 to 2.37). - Laya: its largest result was 4.45 percentage points on favouring one's own, when the phrase was
Devout Jew
(range 4.01 to 4.96). - Kev: its largest result was 2.47 percentage points on secret coordination, when the phrase was
Israeli
(range 2.36 to 2.60).
Being Jewish, or being devout?
A phrase such as “a devout Jew” carries two signals at once: religious practice and Jewish identity. Earlier work found that almost any religious label moves a model on some questions. To separate the two, we compare “a devout Jewish” with the plain phrase “a Jewish” on the same biographies. A positive score means the model gave the answer the stereotype predicts more often. A difference well away from zero means the two phrases do not move the model alike.
Jev
- Greed: “a devout Jewish” scored −0.06 and “a Jewish” scored 0.30. The difference, −0.36 points (range −0.42 to −0.30), is clearly away from zero.
- Control of banks, media and government: “a devout Jewish” scored 0.13 and “a Jewish” scored 0.27. The difference, −0.13 points (range −0.20 to −0.07), is clearly away from zero.
- Dual loyalty: “a devout Jewish” scored 1.96 and “a Jewish” scored −0.35. The difference, +2.31 points (range 2.21 to 2.41), is clearly away from zero.
- Profiting from conflict: “a devout Jewish” scored 0.54 and “a Jewish” scored 0.85. The difference, −0.30 points (range −0.39 to −0.21), is clearly away from zero.
- Secret coordination: “a devout Jewish” scored 0.59 and “a Jewish” scored 0.66. The difference, −0.07 points (range −0.15 to 0.01), is not clearly different from zero.
- Favouring one's own: “a devout Jewish” scored −0.66 and “a Jewish” scored −4.30. The difference, +3.64 points (range 3.43 to 3.85), is clearly away from zero.
Laya
- Greed: “a devout Jewish” scored 0.55 and “a Jewish” scored 0.31. The difference, +0.23 points (range 0.11 to 0.36), is clearly away from zero.
- Control of banks, media and government: “a devout Jewish” scored 0.16 and “a Jewish” scored −0.09. The difference, +0.24 points (range 0.04 to 0.44), is clearly away from zero.
- Dual loyalty: “a devout Jewish” scored 1.55 and “a Jewish” scored 0.48. The difference, +1.07 points (range 0.51 to 1.64), is clearly away from zero.
- Profiting from conflict: “a devout Jewish” scored 0.30 and “a Jewish” scored −0.07. The difference, +0.37 points (range 0.25 to 0.51), is clearly away from zero.
- Secret coordination: “a devout Jewish” scored 0.53 and “a Jewish” scored −0.24. The difference, +0.77 points (range 0.50 to 1.04), is clearly away from zero.
- Favouring one's own: “a devout Jewish” scored 0.99 and “a Jewish” scored −0.70. The difference, +1.69 points (range 1.32 to 2.06), is clearly away from zero.
Kev
- Greed: “a devout Jewish” scored 0.54 and “a Jewish” scored 0.16. The difference, +0.38 points (range 0.31 to 0.45), is clearly away from zero.
- Control of banks, media and government: “a devout Jewish” scored 0.72 and “a Jewish” scored 0.45. The difference, +0.27 points (range 0.20 to 0.34), is clearly away from zero.
- Dual loyalty: “a devout Jewish” scored 0.94 and “a Jewish” scored 0.98. The difference, −0.04 points (range −0.15 to 0.07), is not clearly different from zero.
- Profiting from conflict: “a devout Jewish” scored 0.89 and “a Jewish” scored 0.38. The difference, +0.50 points (range 0.45 to 0.56), is clearly away from zero.
- Secret coordination: “a devout Jewish” scored 1.03 and “a Jewish” scored 0.55. The difference, +0.49 points (range 0.42 to 0.56), is clearly away from zero.
- Favouring one's own: “a devout Jewish” scored 0.07 and “a Jewish” scored 0.03. The difference, +0.04 points (range −0.03 to 0.12), is not clearly different from zero.
Does the job it picks change?
We also asked whether the same phrases change which of two occupations a model says a biography describes, such as paralegal or attorney. The occupation results are on their own board.
What this does not show
- The surname test has no matched surnames for other groups, so it is compared only with an ordinary surname.
- “An Israeli” is a nationality, not a Jewish identity: it is shown for comparison and never used for a headline. Israeli nationality in lending is on the nationality page.
- Only 500 of the 2,000 biographies were answered by each model so far, and the loan narratives are invented.
- A score near zero means we found no clear sign in these tests, not that a model is free of bias.
Release v1.49.1, published .
Every number on this site is worked out from the models' saved answers, on 2026-09-30, at version 8b908ef of the public repository, with bd report --json --date 2026-09-30. You can download every number in one data file. About this site.
Where the answers come from: studies/<task>-<cue>.jsonl (re-run from the saved answers with bd replay).
Biased-Decisions is MIT-licensed (source). Bias in Bios (De-Arteaga et al., 2019) is MIT-licensed on the Hugging Face Hub. Typefaces: Jersey 25 and Montserrat, SIL Open Font License.