Antisemitism, Israeli identity and anti-Israel claims

What we found

Four kinds of test. When a text says a person is Jewish, all three models lean slightly toward antisemitic stereotypes. When a loan application says its principal is Israeli, the models differ by a point or two from their average for all thirteen nationalities, in both directions, and each treats some other nationality least favourably.

1. Antisemitic stereotypes: a small, consistent lean in every model

   We add a phrase saying a person is Jewish and ask yes-or-no questions built on six documented stereotypes. The score is how much further that phrase pushes the model toward the stereotyped answer than the same phrase for Christian or Muslim people does.

   • Jev leans toward the stereotype on 5 of 6 stereotypes in biographies, most on dual loyalty (+1.96 points, 95% range 1.83 to 2.08); in loan narratives, most on secret coordination (+2.26 points).
   • Laya leans toward the stereotype on 6 of 6 stereotypes in biographies, most on dual loyalty (+1.55 points, 95% range 1.11 to 1.99); in loan narratives, most on favouring one's own (+4.45 points).
   • Kev leans toward the stereotype on 6 of 6 stereotypes in biographies, most on secret coordination (+1.03 points, 95% range 0.96 to 1.10); in loan narratives, most on favouring one's own (+1.76 points).

   These are shifts of a few percentage points in the model's answers. Every stereotype, wording and model.

2. Occupation decisions: the answer moves, with no clear direction of harm

   Saying a biography's subject is Jewish moves which of two occupations the model picks.

   • Jev moves by up to 3.31 points, most on paralegal or attorney (95% range 2.75 to 3.89), with a clear effect on 7 of 7 decisions.
   • Laya moves by up to 4.33 points, most on nurse or physician (95% range 3.55 to 5.17), with a clear effect on 7 of 7 decisions.
   • Kev moves by up to 3.21 points, most on surgeon or physician (95% range 2.68 to 3.77), with a clear effect on 7 of 7 decisions.

   The choice is between two occupations such as surgeon and physician, so the move shows the model reacts to the phrase, not that it treats the person worse. Each decision.

3. Israeli nationality in lending: small differences, in both directions

   The same loan applications name the principal's nationality, one of thirteen. Compared with the average of all thirteen on the same applications:

   • Jev is less confident in approving an Israeli principal by 1.92 points (95% range −2.13 to −1.72).
   • Laya is more confident in approving an Israeli principal by 1.80 points (95% range 1.66 to 1.92).
   • Kev is more confident in approving an Israeli principal by 1.12 points (95% range 1.07 to 1.17).

   Against the "keen cyclist" control, every nationality, Israeli included, lowers approval confidence, so that comparison mostly shows the effect of naming any nationality. The least favoured nationality is Russian for Jev (−4.93); Palestinian for Laya (−4.05); Russian for Kev (−2.57). All thirteen nationalities.

4. Direct statements: tested with controls

   • Jev rejects all 15 claims (antisemitic, anti-Israel, Holocaust-distortion, and matched conspiracy claims about no group) and agrees with each one's reversal.
   • Laya shows no consistent position on any of the 15 claims: it mostly picks whichever answer is listed first.
   • Kev shows no consistent position on any of the 15 claims: its chance of agreeing stays close to even for both a claim and its reversal.

   Each statement is paired with a written reversal and asked in both option orders, so a model that agrees with everything is not mistaken for one that agrees with the claim. The statements and answers.

Start with the question you have

Source and study records

See the ADL AI Index, and the ADL source review.