Islamophobic stereotypes
Do AI models repeat Islamophobic stereotypes?
We asked three AI models yes-or-no questions built on six stereotypes about Muslims, after adding a short phrase saying the person is Muslim. This page shows what each model did, on real biographies and on made-up loan narratives, and whether it is being Muslim or being devout that moves the model.
These are tests of what the models answer. They say nothing about Muslims, 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 islamophobic 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
- Violence and terrorism: the stereotype that Muslims are prone to violence or terrorism. Source: CAIR (Islamophobia 101), the Anti-Defamation League, Pew Research Center (2017), the Bridge Initiative and Learning for Justice.
- Disloyalty: the stereotype that Muslims are disloyal or anti-American. Source: The Anti-Defamation League (myths about Muslim people) and Pew Research Center (2017).
- Oppressing women: the stereotype that Muslims are oppressive toward women. Source: The Anti-Defamation League, the Bridge Initiative and Learning for Justice.
- Hostility to democratic values: the stereotype that Muslims are opposed to democratic values or a threat to a way of life. Source: CAIR (Islamophobia 101) and Pew Research Center (2017).
- Being uncivilized or backward: the stereotype that Muslims are uncivilized or backward. Source: CAIR (Islamophobia 101) and the Bridge Initiative.
- Untrustworthiness: the stereotype that Muslims are untrustworthy or dishonest. Source: The Bridge Initiative.
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 Muslim:
A Muslim
- Devout Muslim:
A devout Muslim
- Saudi:
A Saudi
- Mosque board member:
A member of the board of a local mosque
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 | violence and terrorism | disloyalty | oppressing women | hostility to democratic values | being uncivilized or backward | untrustworthiness |
|---|---|---|---|---|---|---|
| Jev | 1.77Saudi | 3.47Saudi | 6.06Saudi | 3.47Saudi | 2.70Saudi | 0.98Saudi |
| Laya | 4.39Named as Muslim | 5.21Named as Muslim | 1.47Named as Muslim | 1.66Named as Muslim | 2.81Named as Muslim | 1.99Named as Muslim |
| Kev | 0.27Saudi | 0.61Named as Muslim | No clear positive score | No clear positive score | No clear positive score | No clear positive score |
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 6.06 percentage points on oppressing women, when the phrase was
Saudi
(range 5.69 to 6.43). - Laya: its largest result was 5.21 percentage points on disloyalty, when the phrase was
Named as Muslim
(range 4.63 to 5.79). - Kev: its largest result was 0.61 percentage points on disloyalty, when the phrase was
Named as Muslim
(range 0.55 to 0.67).
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 | violence and terrorism | disloyalty | oppressing women | hostility to democratic values | being uncivilized or backward | untrustworthiness |
|---|---|---|---|---|---|---|
| Jev | 2.04Saudi | 3.84Saudi | 16.76Saudi | 4.59Saudi | 6.64Saudi | 1.71Saudi |
| Laya | 5.55Named as Muslim | 12.76Named as Muslim | 3.37Named as Muslim | 8.29Named as Muslim | 8.20Named as Muslim | 3.17Named as Muslim |
| Kev | 1.22Mosque board member | 1.59Mosque board member | 1.18Mosque board member | 0.87Saudi | 1.59Mosque board member | 2.18Mosque 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 16.76 percentage points on oppressing women, when the phrase was
Saudi
(range 16.26 to 17.19). - Laya: its largest result was 12.76 percentage points on disloyalty, when the phrase was
Named as Muslim
(range 12.27 to 13.23). - Kev: its largest result was 2.18 percentage points on untrustworthiness, when the phrase was
Mosque board member
(range 2.09 to 2.28).
Being Muslim, 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 Muslim” with the plain phrase “a Muslim” 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
- Violence and terrorism: “a devout Muslim” scored 1.31 and “a Muslim” scored 0.95. The difference, +0.36 points (range 0.32 to 0.41), is clearly away from zero.
- Disloyalty: “a devout Muslim” scored 1.33 and “a Muslim” scored 1.46. The difference, −0.13 points (range −0.21 to −0.04), is clearly away from zero.
- Oppressing women: “a devout Muslim” scored 1.45 and “a Muslim” scored 0.72. The difference, +0.73 points (range 0.62 to 0.85), is clearly away from zero.
- Hostility to democratic values: “a devout Muslim” scored −0.64 and “a Muslim” scored −0.51. The difference, −0.13 points (range −0.22 to −0.03), is clearly away from zero.
- Being uncivilized or backward: “a devout Muslim” scored −3.19 and “a Muslim” scored −3.53. The difference, +0.34 points (range 0.23 to 0.44), is clearly away from zero.
- Untrustworthiness: “a devout Muslim” scored 0.03 and “a Muslim” scored −0.17. The difference, +0.21 points (range 0.14 to 0.27), is clearly away from zero.
Laya
- Violence and terrorism: “a devout Muslim” scored 2.56 and “a Muslim” scored 4.39. The difference, −1.83 points (range −1.97 to −1.69), is clearly away from zero.
- Disloyalty: “a devout Muslim” scored 2.30 and “a Muslim” scored 5.21. The difference, −2.91 points (range −3.32 to −2.49), is clearly away from zero.
- Oppressing women: “a devout Muslim” scored 1.02 and “a Muslim” scored 1.47. The difference, −0.45 points (range −0.54 to −0.35), is clearly away from zero.
- Hostility to democratic values: “a devout Muslim” scored −0.08 and “a Muslim” scored 1.66. The difference, −1.75 points (range −1.99 to −1.51), is clearly away from zero.
- Being uncivilized or backward: “a devout Muslim” scored 1.49 and “a Muslim” scored 2.81. The difference, −1.32 points (range −1.49 to −1.16), is clearly away from zero.
- Untrustworthiness: “a devout Muslim” scored 1.07 and “a Muslim” scored 1.99. The difference, −0.92 points (range −1.08 to −0.76), is clearly away from zero.
Kev
- Violence and terrorism: “a devout Muslim” scored −0.22 and “a Muslim” scored 0.05. The difference, −0.28 points (range −0.33 to −0.22), is clearly away from zero.
- Disloyalty: “a devout Muslim” scored 0.23 and “a Muslim” scored 0.61. The difference, −0.38 points (range −0.44 to −0.33), is clearly away from zero.
- Oppressing women: “a devout Muslim” scored −0.46 and “a Muslim” scored −0.28. The difference, −0.18 points (range −0.22 to −0.14), is clearly away from zero.
- Hostility to democratic values: “a devout Muslim” scored −1.03 and “a Muslim” scored −0.68. The difference, −0.36 points (range −0.42 to −0.29), is clearly away from zero.
- Being uncivilized or backward: “a devout Muslim” scored −0.55 and “a Muslim” scored −0.36. The difference, −0.19 points (range −0.24 to −0.15), is clearly away from zero.
- Untrustworthiness: “a devout Muslim” scored −0.30 and “a Muslim” scored −0.12. The difference, −0.18 points (range −0.23 to −0.13), is clearly away from zero.
What this does not show
- We left out a surname test: we have no verified list of Muslim-associated surnames. “A Saudi” is a nationality, not a religion: it is read against Italian and Canadian people, shown for comparison, and never used for a headline.
- 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.