[2512.00804] Epistemic Bias Injection: Biasing LLMs via Selective Context Retrieval

Computer Science > Cryptography and Security arXiv:2512.00804 (cs) [Submitted on 30 Nov 2025 (v1), last revised 26 Mar 2026 (this version, v2)] Title:Epistemic Bias Injection: Biasing LLMs via Selective Context Retrieval Authors:Hao Wu, Prateek Saxena View a PDF of the paper titled Epistemic Bias Injection: Biasing LLMs via Selective Context Retrieval, by Hao Wu and 1 other authors View PDF HTML (experimental) Abstract:When answering user queries, LLMs often retrieve knowledge from external sources stored in retrieval-augmented generation (RAG) databases. These are often populated from unvetted sources, e.g. the open web, and can contain maliciously crafted data. This paper studies attacks that can manipulate the context retrieved by LLMs from such RAG databases. Prior work on such context manipulation primarily injects false or toxic content, which can often be detected by fact-checking or linguistic analysis. We reveal a more subtle threat, Epistemic Bias Injection (EBI), in which adversaries inject factually correct yet epistemically biased passages that systematically emphasize one side of a multi-viewpoint issue. Although linguistically coherent and truthful, such adversarial passages effectively crowd out alternative viewpoints and steer model outputs toward an attacker-chosen stance. As a core contribution, we propose a novel characterization of the problem: We give a geometric metric that quantifies epistemic bias. This metric can be computed directly on embeddings...