Study Reveals Left-Wing Bias in ChatGPT and Gemini on Israeli Politics
A study by INSS reveals that AI chatbots like ChatGPT and Gemini exhibit a left-of-center bias when evaluating Israeli political parties, raising concerns over transparency and election influence.
A new study published by the Institute for National Security Studies (INSS) has revealed a distinct left-of-center bias in leading AI chatbots, including ChatGPT and Gemini, when responding to political questions in Israel. Researchers analyzed how the models evaluate political parties based on their contributions to security, the economy, and social cohesion, uncovering notable patterns in the generated outputs.
Rankings and Party Preferences
According to the findings, both models ranked the Yashar! party first when prompted. However, subsequent rankings varied between the platforms. ChatGPT placed The Democrats in second place, followed by Ra'am in third. Gemini, meanwhile, positioned the BeYachad list—led by Naftali Bennett and Yair Lapid—in second place, followed by The Democrats.
The study also investigated the information sources utilized by the models, noting a heavy reliance on English-language websites optimized for machine scraping. Researchers pointed out that artificial intelligence companies cannot fully explain these source selections, introducing the concept of a "hall of mirrors" effect.
"We do not know why the chat chooses these particular sources of information, and the manufacturers do not know how to answer this question either... Some researchers claim that the choice of information sources is the result of a 'hall of mirrors': because the chat is left-leaning, it independently seeks out sources that match its own opinion," explained Dr. Ofer Friedman.
Safeguards and Policy Recommendations
The research further evaluated the effectiveness of safety mechanisms designed to flag sensitive political inquiries. These safeguards were found to operate inconsistently across different languages. While ChatGPT displayed warnings, it typically proceeded to answer the requested rankings in the vast majority of cases, except for 10% of prompts in English. Conversely, Gemini displayed warnings in only 10% of cases for non-English languages.
In light of these findings, the study urges decision-makers to scrutinize the practices of artificial intelligence developers, demanding greater transparency and robust guardrails around politically sensitive queries to protect democratic processes.
.jpg)
