The Palestine-Israel Conflict in AI Code: Enhancing Geopolitical Narrative Allegiance in Fine-Tuned LLMs

<p><span>Since the launch of OpenAI’s ChatGPT, artificial intelligence (AI), and large language models (LLMs) in particular, have initiated a shift in how users access knowledge, away from traditional keyword-based queries on search engines and toward natural language interaction. This development has positioned LLMs as a major tool in the user toolbox across the globe, including in the Global South. However, the literature has repeatedly shown that LLMs tend to exhibit biases in favor of Western culture and worldviews. This tendency largely stems from the fact that most LLMs are trained on data that disproportionately originates from Western contexts and is subject to guardrails and instructions based on Western values, a structural condition that has also been shown to shape global media outputs. As argued by Gramscian theories of hegemony and by the Propaganda Model developed by Herman and Chomsky, such bias is not anomalous but rather a predictable outcome of Western dominance in political, cultural, and economic domains. Rather than surrendering to another wave of cultural domination driven by globally deployed LLMs that are culturally misaligned with many non-Western societies, several initiatives have emerged from the Global South to optimize and fine-tune popular open-weight models to better align them with regional cultural values.</span></p> <p><span>This study puts one such initiative to the test by applying a geopolitical stress-test approach to a highly polarizing issue, namely the Israeli–Palestinian conflict. To control for the effect of simple hedging mechanisms, derivative experimental controls were developed in which a base-architecture model developed by Google (Gemma2), operationalized as a proxy for the hegemonic sphere, was paired with a regionally aligned, fine-tuned derivative model developed in Qatar (Fanar1). The study measures agreement with paired Israeli-centric and Palestinian-centric frames across several narrative zones. Quantitative analysis reveals a strong statistical divergence, indicating that the derivative model performs an act of algorithmic counter-hegemony. This divergence is most pronounced in relation to legal and infrastructural issues, such as settlements and home demolitions. By contrast, notable convergence toward neutrality is observed in the security frame that involves active violence as resistance to Israeli occupation. These findings suggest that while political filters may be contested through fine-tuning, architectural safety layers tend to be more persistent. Although the results demonstrate that hegemonic narratives can, to some extent, be countered through optimization and fine-tuning, they also point to the risk of algorithmic Balkanization, whereby discursive space may fragment into mutually incompatible local worlds.</span></p>