Abstract
<p>Large Language Models (LLMs) are rapidly replacing traditional search engines as people’s primary gateways to information. Yet, despite their growing significance and the increasingly central role of LLMs in political communication and participation, legal frameworks provide limited opportunities for transparency, let alone changes, over how politically sensitive prompts are handled. At the same time, scholarly debates about LLM objectivity remain fixated on statistical formulas for measuring political bias that presuppose a neutral ‘centre’ for contested issues (e.g. abortion), the deviation from which can be measured following traditional mental maps of bias analysis; an assumption, however, that lacks foundation in political and legal philosophy. This paper proposes an epistemic shift in how neutrality is framed and dealt with in LLM research and practice. First, we systematically critique the dominant ‘political bias’ paradigm, showing how scholars as well as government and industry actors frame neutrality in ways that entrench methodological limitations. Then, we conduct a multi-stage, mixed-methods pilot study to illuminate aspects of GPT-5’s web search that are core to the processes of information retrieval. To do so, we prompted GPT-5 API with a corpus of binary and politically loaded prompts focusing on the human rights stances of notable political figures (US, UK, EU). For each prompt, we retrieve all URLs (n= 22,500), search queries, and reasoning traces, paying close attention to the logic and parameters disclosed by GPT-5’s web search. After that, we deployed qualitative and quantitative methods to map how the model assesses the credibility of sources and to measure discrepancies between domain retrieval and citation frequencies, respectively. Departing from output-focused bias studies, our work reveals the infrastructural terrain where ‘LLM neutrality’ materialises. In particular, we: 1) identify misconceptions about how political bias is manifested and understood in practice; 2) pinpoint architectural details, operational gaps, and double standards in GPT-5’s web search functions; and 3) offer interdisciplinary recommendations for policymakers, practitioners, and researchers in the field.</p>