In press
Horizon Translation for Artificial General Intelligence: Definitions, Framework, and Metrics
Abstract
As AI systems approach and exceed human-level performance, the central difficulty shifts from whether a system can solve a problem to whether humans can still follow, verify and act on how it did so. This paper introduces horizon translation: the faithful, verifiable transfer of justified understanding across the gap between two minds’ cognitive horizons. We define the cognitive horizon and the co-horizon, the region where human and AI understanding overlap and where oversight and collaboration can work. We then propose a framework with three translation layers (explanatory, verificatory and predictive), four directions of translation, and five translation modes that range from narrative explanation to machine-verifiable proof. We identify six failure modes: lossy, persuasive, lagging, illusory, overloading and domain-mismatched translation. We argue that persuasive translation is the most dangerous, because it raises reviewer approval without raising accuracy and so resembles success. We propose six metrics, namely translation fidelity, verification yield, calibration gap, predictive accuracy, time to understanding and persuasion index. We also give a first-pass fidelity score weighted by stakes. We relate the concept to interpretability, explainability, scalable oversight, alignment and calibration, and outline four scenarios for how capability and translation fidelity may evolve together or apart. We close with eight open research questions, working responses to each, and five recommendations. The central one is to treat translation fidelity as a first-class property alongside capability. The metrics are starting points, not validated instruments.
Keywords
horizon translationartificial general intelligencescalable oversightco-horizontranslation fidelitypersuasive translationinterpretabilitycalibrationhuman–AI alignmentAI safety