Epistemic Filtering and Collective Hallucination: A Jury Theorem for Confidence-Calibrated Agents

A new research paper by Jonas Karge, titled 'Epistemic Filtering and Collective Hallucination: A Jury Theorem for Confidence-Calibrated Agents,' introduces a formal mathematical framework for improving the accuracy of collective AI decision-making. The core innovation is moving beyond classical voting models where all agents must participate, to a system where AI agents can learn their own competence over time and say 'I don't know.' This 'epistemic filtering' process involves a calibration phase where agents update beliefs about their reliability, followed by a confidence gate that determines whether they vote or abstain on a final decision. The work directly addresses a key challenge in AI safety: mitigating hallucinations when multiple large language models (LLMs) or AI agents work together.

The paper derives a non-asymptotic lower bound on a group's success probability, proving that this selective participation generalizes the asymptotic guarantees of the famous Condorcet Jury Theorem to a sequential, confidence-gated setting. Empirically, the theoretical bounds are validated through extensive Monte Carlo simulations. While the results are general, the author highlights a direct application to AI safety, outlining how the framework can be used to design systems that reduce collective hallucinations in ensembles of LLMs. This provides a principled, mathematical basis for building more reliable multi-agent AI systems where not every model has to answer every question, potentially leading to more trustworthy and accurate collective outputs.