Mine over Yours: How Authorship Biases Evaluation in Generative Information Retrieval
Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2026)
Generative Information RetrievalInteractive RetrievalAuthorship BiasIKEA EffectInformation TrustQuality JudgmentHuman-AI Interaction
Does an answer seem better because you helped produce it?
Conversational search asks us to do more than enter a query. We ask follow-up questions, refine the direction, and help shape the answer. This study asks whether that involvement changes how we judge the final information, even when someone else receives equivalent content.
Compare information across conversational sessions
In a mixed-methods experiment with 28 participants, people used an AI system to retrieve and develop information through conversation. They then evaluated information from their own sessions and equivalent AI-curated information associated with other users. The study examined quality judgments, trust, and the role of the effort invested in the interaction.
Ownership changed quality judgments more than trust
Participants rated self-obtained information more highly for quality, while trust ratings remained similar across conditions. Greater effort amplified the preference for one’s own results. Awareness that AI can hallucinate did not eliminate this bias in evaluating and selecting information.
Engagement can make critical evaluation harder
A conversational system can encourage involvement while also making its output feel like our own work. The design challenge is to support that involvement without letting ownership stand in for evidence of quality. This study identifies a bias in the tested retrieval setting; it does not show that information from one’s own session is always wrong or less trustworthy.
Read the original abstract
Generative information retrieval (GenIR) enables users to obtain information through iterative LLM interaction rather than merely retrieving existing documents. We examine whether this active involvement biases users’ evaluation of AI-crafted information—specifically, whether users judge information they obtained through their own interaction more favorably than equivalent information retrieved from others. In a mixed-methods experiment (N=28, 2x2 within-subjects), participants used an AI system to iteratively retrieve and craft informational content, then evaluated their results against equivalent AI-curated information retrieved by others. Results reveal a selective authorship bias: participants significantly overrated self-obtained information on quality, but maintained uniform trust ratings across conditions, reflecting hallucination awareness that nonetheless failed to correct quality-driven selection behavior. Higher effort further amplified this selection bias despite the presence of information conflicts. Since the iterative interaction that triggers this bias is inherent to GenIR, these findings point to a structural challenge requiring system-level safeguards for critical information evaluation.