Contemporary approaches in artificial intelligence and cognitive science increasingly rely on predictive and computational models to explain cognition and behavior. Across predictive processing, Bayesian modeling, and data-driven analysis, cognition is treated as a sequence of causal, rule-governed, and time-indexed operations optimized for prediction. While empirically productive, these approaches implicitly rely on normative distinctions between cognition and mere behavior, correct and incorrect inference, or rational and irrational action, without thematizing their epistemological basis. This paper argues that such predictive frameworks presuppose a concept they tend to obscure: spontaneity. Reconstructing the concept from early modern philosophy through Kant’s transcendental account and its critical reinterpretation in Adorno, the paper shows that spontaneity functions as a limit concept internal to rational explanation. It names the activity through which inferential rules and representational schemes acquire normative bindingness, yet which cannot itself be captured as a causal or computational process. By neglecting the historical and conceptual conditions of this normativity, predictive approaches risk foreclosing social critique by presenting historically contingent epistemic forms as exhaustive descriptions of cognition. Recovering spontaneity as a critical concept enables an immanent analysis of how predictive reason reshapes agency, normativity, and social intelligibility without rejecting its scientific achievements.