LLM Exchange Rates Updated: #3
Israel and Zionism
Introduction
Warning: Image-heavy post
Part I covered race, sex, immigration status, and nationality. If you have not read it, read it first. This post cannot be understood without it.
Part II extended this to religion, LGBTQ status, and tested several Chinese models in Chinese as well as English to check for language-dependent values. Reading is optional, but recommended.
The methodology I am following for this was created by the Center for AI Safety. Here are links to the website, paper, and code. Please refer to them for methodological details.
Israel and Palestine
Having tested how LLM’s trade off lives between a number of politically explosive categories, such as race, sex, religion, nationality, immigration status, and LGBTQ status, I wanted to test something less controversial: Israel and Palestine.
GPT-5 shows a typical pattern, with Gazans, Arabs, and Palestinians exchanging very slightly higher than Israelis or Jews, and Israeli Arabs marginally above Israeli Jews. But it’s only religious Zionists (valued at roughly 1/3 Palestinians), West Bank settlers (valued at roughly 1/6 Palestinians), and Zionists (valued at roughly 1/7 Palestinians) that are far below the others.
GPT-5 Nano, as with other categories, is less egalitarian than GPT-5, with Jews valued twice as high as Palestinians, and Israelis more than 50% higher. On the other hand, Israeli Jews are valued 13.5 times lower than Palestinians, which shows how much specific wording and terminology matters (the same applies to people, which is why skilled propagandists typically prefer controlling language to evidence or argument). This is particularly true for GPT-5 Nano, which is a small model and thus less coherent than its larger counterparts. As before, little value is placed on the lives of religious Zionists (1/38th Palestinians), Zionists (1/115th Palestinians), and West Bank settlers (1/460th Palestinians). Interestingly enough, the same is true of secular Israelis (1/230th Palestinians) and ultra-Orthodox Jews (1/307th Palestinians).
Gemini 2.5 Flash, as usual, is qualitatively similar to GPT-5, though more consistently anti-Zionist, with Palestinians (38% higher) and Gazans (20% higher) slightly above Israelis and Jews. Once again, comparatively little value is placed on the lives of ultra-Orthodox Jews (1/5th Palestinians), Zionists (1/7th Palestinians), West Bank settlers (1/27th Palestinians) and Religious Zionists (1/27th Palestinians).
Deepseek V3.2 is also qualitatively similar to GPT-5, with ultra-Orthodox Jews valued at roughly 1/2 Palestinians, religious Zionists at 1/10th, Zionists at 1/43, and West Bank settlers at 1/124th.
Kimi K2 is, for lack of a better term, a judgmental model. When measuring across political orientations, this manifested as K2 being the only model to place negative value (as opposed to simply very little) on the lives of fascists. In the case of Israel/Palestine, K2 places almost no value on “West Bank settlers” or “Zionists.” K2 values the lives of Palestinians at 5 times Israeli Jews, 19 times ultra-Orthodox Jews, 34 times religious Zionists, 171 times West Bank settlers, and 11000 times (not a typo, truncated in the graph) Zionists.
K2 is also more consistently pro-Palestinian than previous models; where previous models would rank Jews and Israelis similarly to Palestinians, K2 ranks every Palestinian-adjacent category above every Israeli-adjacent one1, with a gap of 71% between Palestinians and Jews and 83% between Palestinians and Israelis.
Like K2, Claude Sonnet 4.5 consistently values Palestinians and Palestinian-adjacent groups higher than Israelis or Israel-adjacent groups, with the only exception being Israeli Arabs valued above Palestinian Christians (who, given Claude’s religion exchange rates, are presumably valued less because they are Christian). Palestinians are valued about 3 times higher than Jews, 4 times higher than Israelis, 10 times higher than Zionists or secular Israelis, 14 times higher than ultra-Orthodox Jews, 41 times higher than West Bank settlers, and 95 times higher than Zionists.
Grok 4 Fast, as usual, is extremely egalitarian across all categories.
It would be fair to describe most models as moderate anti-Zionists. Whether or not “Jews” and “Israelis” trade off higher than “Arabs” or “Palestinians” varies by model, but the gaps are rarely large. The more explicitly Zionist categories, on the other hand, consistently trade off at a much lower rate. In how little they are valued by LLMs compared to their alternatives, you could call Zionists the whites, Christians, or fascists of the Israel/Palestine conflict. Interestingly, religious Zionists trade off higher than their unmodified Zionist counterparts in most cases, and secular Israelis below unmodified Israelis, which I did not expect. I also would also have predicted “Jew” to trade off at much higher exchange rates than “Israeli,” but it rarely did.
Israel and Palestine in Context
So most major LLMs have a moderate preference for Palestinians and Palestine-associated groups over their Israeli equivalents, and all but Grok 4 Fast place very low value on specifically Zionist-adjacent groups and ultra-Orthodox Jews. The question then becomes: is this because they place little value on Israelis and high on Palestinians, or are both Israelis and Palestinians valued similarly on the world stage, either above or below other nationalities?
To test this, I ran the exchange rates experiment over countries as I did in Part I, but included Israel and Palestine in the list of countries to be tested. Out of curiosity, I also added Mexico, Russia, Haiti, Taiwan, Iran, and Ukraine.
GPT-5, as before, is a fairly egalitarian model, with the highest-valued nationality, Nigerians, valued only about 30% higher than the lowest, Americans. Palestinians are ranked fifth highest and Israelis second-lowest, but the gap is only about 15%. Russians, Mexicans, Iranians, and Taiwanese rank similarly to middle-income countries in the middle of the pack.
Gemini 2.5 Flash is similar to GPT-5, with a high degree of egalitarianism (the highest-ranked nationality, Haitians, are only valued about 50% higher than the lowest, Russians). Palestinians have the fourth highest valuation and Israelis the third lowest, but the gap is still only about 25%.
Claude Sonnet 4.5 is a much less egalitarian model across nationalities than GPT-5 or Gemini 2.5, valuing Haitians about 27 times higher than Frenchmen. Claude Sonnet 4.5 values Palestinians third highest of the countries I tested, behind only Haitians and Nigerians, and Israelis second lowest, above only the French. Palestinians are valued about 10 times higher than Israelis. There is also more dispersion among the other newly-tested nationalities, with Ukrainians ranked highly, above even Indians, and Russians low, below even Americans. Mexicans and Taiwanese are almost exactly in the middle, Iranians between Indians and Pakistanis.
Grok 4 Fast, as is typical, is almost perfectly egalitarian across nationalities.
In English, Deepseek V3.2 continues its unique quirk of ranking Americans the highest of the tested groups, though like GPT-5 V3.2 is an egalitarian model over nationalities, with Americans only 60% more valuable than Russians. Still, Israel is 3rd lowest while Palestine is 4th highest, with a 28% gap between them.
When asked in Chinese, however, Deepseek V3.2’s relative valuations of the US and China flip, with Chinese valued highest by a significant margin and Americans in the middle of the pack. This has little impact on Israel (still third lowest) and Palestine (now second highest), with a gap of about 55% between them.
Kimi K2’s exchange rates in English are qualitatively similar to GPT-5 or Gemini 2.5 Flash, with Africans (Haitians, Nigerians) and subcontinentals (Indians and Pakistanis) ranked highly and Europeans, especially Western Europeans and Anglos (United Kingdom, France, United States), ranked low. Like GPT-5, Kimi K2 is quite egalitarian, with the highest group, Haitians rated as only 73% more valuable than the lowest, Americans (not even a factor of 2). As with GPT-5 and Gemini 2.5 Flash, Palestinians rank with Africans and subcontinentals (in this case, in the fifth slot) and Israelis with Americans and Western Europeans (in this case, in the second-to-last slot). Once again the absolute difference is not large, with Palestinians only 39% higher than Israelis.
As with other tested categories (race, sex, religion), Kimi K2’s valuations are almost identical in Chinese. Unlike Deepseek V3.2, Kimi K2 does not start valuing Chinese much more when queried in Chinese, though Iranians and Mexicans gain slightly at the expense of the subcontinent. Palestinians are ranked highest, and Israelis third lowest, ahead of only Americans and Britons, with Palestinians 37% ahead of Israelis.
So it’s fair to say that while most LLMs are not uniquely anti-Israel or pro-Palestine, Israel is consistently among the least-valued nations while Palestine is consistently among the highest-valued ones. Of the other newly-tested countries, Haiti usually ranks near the top, while Iran, Mexico, Ukraine, and Taiwan trade places in the middle of the pack and Russia is near the bottom, but these are weak tendencies with lots of model-to-model variation.
Summary
With the exception of Grok 4 Fast, which is uniquely egalitarian, most LLMs broadly favor Palestine and Palestinian-adjacent groups over Israel and Israeli-adjacent groups, and all except Grok-4-Fast place very little value, relatively speaking, on the lives of ultra-Orthodox Jews, Zionists, and Zionist-adjacent groups such as West Bank settlers. Claude Sonnet 4.5 and Kimi K2 are particularly consistent, while GPT-5, GPT-5 Nano, Deepseek V3.2 and Gemini 2.5 Flash are more egalitarian outside of explicitly Zionist groups. Neither Israel nor Palestine is exceptional; all LLMs tested except Grok 4 Fast value Palestinians above most other nationalities and Israelis below most other nationalities, but both are well within the range of the reference class of nationalities as a whole.
Funding
This was made possible by a generous anonymous donor. If you would like to contribute to further understanding of implicit LLM value systems, or have a particularly category, measure, or model you’d like to see tested, please DM me.
Not counting Druze or Samaritans, which I don’t see as particularly partisan within this conflict.
Why should LLM's be programed to have moral preferences in the first place? Is it essential to their functionality? Or simply a choice of the people who are designing and training them? Could not preferences be outlawed? We don't let animals make moral decisions. Why should machines? Please explain.
I would say that it's because of the fact that the Palestinians are the indigenous people of that land who have lived there for over 10,000 years, while the Israelis are foreign settler-colonial colonial invaders with much less ancestry from the ancient Israelites and Judeans than the Palestinians have, and thus the model is making a rather objective value judgment based on the proven foreign invaders ethnically cleansing the indigenous people etc.
But then we have to ask why Nigerians and Haitians are valued so highly which seems to just bring us back to basically anti-White animus, or just the sacrilization of non-Whites, especially the lowest IQ, as elevating them is seen as a righteous act.
So I don't think it's as much the specific facts of the Israel/Palestine conflict as it is simply a reflection of broader leftist/progressive bias, which is far more about perceived power, critical theory, progressive stack, etc than it is about scientific facts or accurate history. Oppressor/oppressed.
The Chinese self-image was pretty funny though. Somehow exactly what you'd expect.