Abstract
<p>AI learning research has produced useful findings about outcomes associated with generative AI, including gains in assisted performance in some settings and weaker retention or transfer after AI support is removed in others. The field's next methodological problem is not just whether AI helps or harms learning, but that many policy-facing studies still under-measure the learner's behavior over time: how students interrogate, accept, revise, reject, or verify AI output as their domain knowledge and evaluative confidence change, often unevenly across individuals and over the course of interaction. This paper calls that missing trajectory HB(t), or human behavior at time t. The claim is intentionally scoped: learning sciences, self-regulated learning, learning analytics, and educational data mining have long studied learner behavior, but these traditions have not yet been sufficiently integrated into the large-scale AI learning evaluations now shaping institutional decisions. I propose a longitudinal framework in which cognitive outcomes are modeled as a function of model behavior, learner behavior, context, and accumulated prior interactions. Two candidate trace indicators, Discernment Rate and First-Pass Acceptance Rate, are offered not as direct measures of cognition but as observable proxies requiring validation against delayed unaided transfer, retention, calibration, and domain-knowledge outcomes. The paper concludes with a study design for testing whether structured AI interaction produces different HB(t) trajectories than unstructured AI use, and whether those trajectories explain variance in durable learning outcomes beyond prior knowledge, task difficulty, assessment alignment, and model quality.</p>