Neural retrievers have a documented tendency to rank LLM-generated passages above human-written ones that are semantically similar. According to the abstract of a new arXiv paper, this source bias threatens the reliability of modern information access systems.

The paper's title points to a specific mechanism: relevance supervision. The central claim is that training with relevance labels can teach retrievers to prefer LLM-generated text, suggesting the bias may be learned rather than intrinsic to the models themselves.

Because this is a single preprint, the findings should be treated as preliminary. The abstract does not include experimental details, but it frames the issue as a growing concern for systems that depend on neural retrieval.