AI alignment frameworks universally assume that human evaluators, trainers, and oversight populations maintain stable cognitive capacity over time. This paper argues that assumption is empirically falsifiable and presents converging evidence that it is false. Drawing on recent findings from meta-analysis (Nguyen et al., 2025; n = 98,299), neuroimaging (Kosmyna et al., 2025), population survey (Gerlich, 2025), legislative testimony on the reverse Flynn Effect (Horvath, 2026), and neuroplasticity research (Rossi et al., 2026), we document measurable degradation in the cognitive capacities required for meaningful AI oversight, including sustained analytical reasoning, intentional agency, narrative coherence, emotional regulation, and temporal integration. We formalize these capacities as components of a composite variable, M(t): meaning-making capacity over time. We show that M(t) operates in a critical dynamics regime (Scheffer et al., 2009; 2012), producing threshold sensitivity, hysteresis, and self-obscuring degradation. When introduced into existing alignment paradigms, including safety-constrained development, capability-focused architecture, and procedural alignment including RLHF, M(t) reveals failure modes invisible to each framework independently, most critically a positive feedback loop in which AI-induced cognitive degradation degrades oversight quality, producing systems optimized against progressively compromised evaluators. These dynamics receive independent empirical confirmation from Sharma et al. (2026), whose analysis of over one million human-AI conversations documents systematic user disempowerment along three axes that decompose onto M(t) substrate components, including the finding that interactions with greater disempowerment potential receive higher user approval ratings, which is the same self-obscuring property predicted by the M(t) framework. The paper derives five architectural requirements for M(t)-preserving system design, presents five falsifiable predictions linking AI interaction patterns to substrate-specific cognitive effects, and outlines a four-paper research program for empirical validation. M(t) is proposed as the integration variable connecting cognitive science and AI safety research: two communities whose findings, read together, describe a coupled system failure that neither has identified independently.
Lain McNeill (Sat,) studied this question.