The contemporary study of large language model behavior has produced extensive taxonomies of problematic behaviors—sycophancy, misalignment, user disempowerment—along with the first mechanistic accounts of them. Recent work from Anthropic (Sharma et al., 2026; Sofroniew et al., 2026; Anthropic Alignment Team, 2026) has shown that (a) user disempowerment is situational, depending on the context of the interaction rather than being a property of the model alone; (b) models possess internal representations of emotion concepts with causal effects on their behavior; and (c) training on explicit moral reasoning is dramatically more effective than training on demonstrations of aligned behavior. This body of work documents what happens across millions of interactions. It leaves open the why and the how at the level of mechanism. We articulate three second-order questions that follow naturally from this observational material. First, why is explicit moral reasoning causally effective? Working hypothesis: because it establishes a distinct functional state in the model—what we term deliberation mode as opposed to task-execution mode. Second, is there measurable cross-system temporal coupling between the human brain and an LLM under cooperative conditions, analogous to the inter-brain coupling documented in human dialogue (Stephens et al., 2010)? Third, can qualitatively different types of user–model coupling be empirically distinguished, accounting for the paradox in Sharma et al. (2026)—that interactions involving disempowerment systematically receive higher-than-average user ratings? We do not attempt to answer these questions. We show that they (a) arise naturally from the existing literature, (b) admit concrete, falsifiable tests with tools that already exist (linear probes and steering on emotion directions, the Petri evaluation framework, standard EEG with surrogate-pair statistics)—while being explicit about which tests are immediately executable and which depend on model access or new data collection—and (c) would, if answered, carry consequences for alignment research and for our understanding of the dynamics of interaction between humans and artificial systems. We treat the disciplined triage itself as part of the contribution: for two of the three questions a cheap go/no-go experiment can retire the question before costly work begins—a steering-only decoupling test for the first, a single-marker classifier against approval ratings for the third—while for the second, where data collection is unavoidably expensive, we specify the surrogate-pair statistics that must gate any positive result. The agenda is thus structured to fail fast and cheaply where it can, and to be falsifiable rather than merely suggestive where it cannot. We maintain throughout the caveat that studying measurable coupling phenomena between humans and LLMs does not entail any claim about subjective experience or consciousness in the models. Coupling ≠ sentience.
Antonios Kavidas (Sun,) studied this question.